Ensemble Forecasting of Power Quality Parameters
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Electrical Engineering. The day in electrical engineering: signals, systems and the hardware that carries them.
Rosa: We're looking at the paper: "Ensemble Forecasting of Power Quality Parameters".
Dev: Here's the gist of the paper. Ensemble forecasting is confirmed as a robust and scalable approach for long-term PQ prediction in transmission systems The results show that ensemble forecasts consistently outperform individual models in terms of accuracy and robustness, achieving significant improvements over seasonal naive benchmarks and the best-performing single models Pruning redundant reasoning tokens The growing integration of power electronicbased technologies has increased the necessity of power quality (PQ) monitoring in transmission systems Medium- to long-term forecasting can enhance the value of these datasets by enabling proactive asset management and trend detection, despite challenges related to data heterogeneity and seasonality This paper systematically evaluates individual and ensemble forecasting approaches for PQ parameters in transmission systems More than 700 weekly time series from measurement campaigns in Germany and Estonia are analysed to assess various models and aggregation strategies within a structured ensemble framework How the training signal is built The dataset originates from two measurement campaigns within the transmission system of a German and Estonian TSO It contains 14 German and 13 Estonian measurements sites spanning different voltage levels: 110 kV (17 sites), 220 kV (2 sites), 330 kV (2 sites), and 380 kV (6 sites) The monitored PQ parameters include different voltage quality parameters: long-term flicker severity (Uplt), voltage unbalance (UNB), total harmonic distortion of voltage (Uthd) and individual harmonic voltages (U03, U05, U07, …, U15) recorded at a 10-minute aggregation interval The original 10-min measurements are first aggregated to weekly values by computing the 95th percentile for each full calendar week (Monday to Sunday) A week is considered valid if at least 95 % of the 10-min measurements are available, allowing for up to 8.5 h of missing data per week in order to increase data availability Missing weeks or gaps within the resulting weekly time series occur frequently, for example due to maintenance outages To maximize the number of usable time series, a up to 20 % of missing weeks is allowed Gaps are filled using the most recent weekly 95th percentile available within the last 10 weeks The weekly values are then normalized by relating them to their respective planning levels, which provides the respective utilization Where the gains come from The performance evaluation procedure maintains consistency by using a 105-week training period and a 52-week forecast horizon for all time series and models Accuracy is assessed in three steps: 1.
Rosa: Calculation of accuracy (MAE and sMAPE) over the entire horizon for each model and time series 2.
Dev: Determination of the rank R(sMAPE) for each model per time series 3.
Rosa: Calculation of the mean accuracy (MAE and sMAPE) and mean rank (R(sMAPE)) for each model across all 716 series The construction of the not equal weighted ensemble forecasts y i is defined by a linear combination of M individual model forecasts y i: y i = ∑ w i ⋅ f(y 1, y 2, …, y M) To ensure that models with higher accuracy (lower error or lower rank) receive greater weight, the weights w i are defined by the reciprocal of a performance metric phi i normalized across all models: w i = phi i / ∑ phi j - 1 Ensemble Model Performance The analysis of 988 ensemble models shows significant improvement over individual models The top 100 models outperforming the best individual model (STL-ARIMA, sMAPE = 18.22 %) The D28 ensemble configuration emerges as the overall winner, where the Median combination achieved the lowest sMAPE of 17.68 % and a Benchmark Ratio of 0.847 Increasing the number of individual models in ensembles consistently reduces both the average sMAPE, improving accuracy, and the performance variance, enhancing robustness Comparison of the Best Individual and Ensemble Model The relative sMAPE improvement shows that the best ensemble outperforms the best individual model in a majority of cases, achieving lower sMAPE in 62 % of all comparisons When the ensemble is superior, the gains are systematic and practically relevant, with a median relative improvement of about 4 % By comparing individual versus ensemble sMAPE, it can be seen that the ensemble stabilizes performance, preventing the extreme forecast failures often associated with selecting a single, sub-optimal model Ensemble methods effectively mitigate this modelselection risk by consistently improving forecast accuracy and reducing error variability The results further indicate that ensemble performance depends primarily on the diversity of the individual models rather than on the complexity of the aggregation method, as simple mean or median combinations already yield robust and reliable improvements Consequently, ensemble forecasting offers a scalable and computationally efficient framework for transmission system operators to enhance trend identification and support infrastructure planning Future research will focus on an application to distribution systems, extending it toward probabilistic forecasting, and developing asymmetric accuracy measures that prioritize forecast accuracy near power quality limits where prediction errors have the greatest operational impact ACKNOWLEDGMENT This work was supported by the German transmission system operator TransnetBW, the Estonian transmission system operator Elering AS, and the European Union, and was co-funded by the Estonian Ministry of Education and Science (Project No.
Title and authors: Rosa: So we’re looking at this paper now, "Ensemble Forecasting of Power Quality Parameters," and it’s about using multiple models to predict power quality in transmission systems <ref:2603.02706#pg1>. It looks like they are tackling a big problem because integrating more power electronics means we need better monitoring than just checking compliance <ref:2603.02706#pg1>.
Dev: Right, and it’s not just about checking boxes anymore; medium- to long-term forecasting can actually help with proactive asset management and spotting trends before things go wrong despite the data being messy <ref:2603.02706#pg1>. The authors are systematically comparing individual models against structured ensemble frameworks using over seven hundred weekly time series from Germany and Estonia <ref:2603.02706#pg1>.
Taro: I’m thinking about those data sources, looking at both German and Estonian TSOs across different voltage levels, that sounds like a lot of heterogeneity to handle <ref:2603.02706#pg1>. If the data is that mixed up, how reliable are these forecasts going to be in the real world <ref:2603.02706#pg1>?.
Rosa: Exactly, and the paper explains how they handle those issues by first aggregating ten-minute measurements into weekly values using a 95th percentile calculation for each full calendar week <ref:2603.02706#pg2>. They even allow up to twenty percent of missing weeks, filling the gaps with the most recent available data from the last ten weeks <ref:2603.02706#pg2>.
Dev: And they normalize those weekly values by relating them to specific planning levels—like VDE-AR-N four thousand one hundred twenty in Germany or whatever level Estonia uses based on IEC TR sixty-one thousand-three-six <ref:2603.02706#pg2>. That normalization step is key because it tells you the actual utilization of the system <ref:2603.02706#pg2>.
Taro: So, before we even look at the models, they’ve got this solid way of cleaning and preparing the data so that it makes sense for long-term prediction <ref:2603.02706#pg1>. What about the forecasting models themselves? Are they just using standard statistical tools or something more advanced <ref:2603.02706#pg1>?.
Rosa: Well, they start with baseline methods like Naive and Drift, and then move into exponential smoothing like Holt Winters and ARIMA models <ref:2603.02706#pg3>. They also mention Prophet as an additive regression framework that handles trend, seasonality, and holiday components <ref:2603.02706#pg3>.
Dev: But they point out a limitation right there; machine and deep learning models are often hindered in long-term forecasting because they rely on specific trend and periodicity structures <ref:2603.02706#pg2>. Plus, those "black box" approaches don't offer much explainability, which is a concern for system operators <ref:2603.02706#pg2>.
Taro: So the authors are setting up this structured ensemble framework where they focus only on ensembles made from separate, noninterfering models <ref:2603.02706#pg3>. That’s a specific way they’re narrowing their scope, focusing on combining forecasts from different sources <ref:2603.02706#pg3>.
Rosa: They test all two hundred forty-seven possible configurations of combinations using the eight individual forecasting models listed in Table I <ref:2603.02706#pg3>. This means they are exploring how different model sizes—from just two models up to all eight—affect the final result <ref:2603.02706#pg3>.
Dev: And the combination function is defined by a linear combination, where the weights for each individual model are set based on its performance metric phi, normalized across all models <ref:2603.02706#pg1>. It ensures that models with lower error or lower rank get more weight in the final forecast <ref:2603.02706#pg1>.
Taro: That sounds like a smart way to weigh the models, making sure the better performers actually influence the final output, rather than just averaging them out <ref:2603.02706#pg1>. How do we know if this combination method is actually working better than just picking one good model <ref:2603.02706#pg8>?.
Rosa: The results show that the ensemble forecasts consistently outperform individual models in terms of both accuracy and robustness <ref:2603.02706#pg1>. They achieve significant improvements over seasonal naive benchmarks and even the best single models <ref:2603.02706#pg1>.
Dev: Specifically, the analysis of nine hundred eighty-eight ensemble models shows a significant improvement over individual models; the top one hundred models outperform the best individual model, which was STL-ARIMA with an sMAPE of eighteen point two two percent <ref:2603.02706#pg7>.
Taro: That means even if you pick the single best model on its own, you’re still losing out when you combine them correctly <ref:2603.02706#pg8>. What did the best ensemble configuration actually achieve <ref:2603.02706#pg7>?.
Rosa: The D28 ensemble configuration turned out to be the overall winner, where the Median combination achieved a very low sMAPE of seventeen point six eight percent and a Benchmark Ratio of zero point eight four seven <ref:2603.02706#pg7>. That’s quite low error for long-term forecasting <ref:2603.02706#pg8>.
Dev: That ratio, that Benchmark Ratio of zero point eight four seven against the best individual model's sMAPE of eighteen point two two percent, it shows how much stability the ensemble adds to the prediction <ref:2603.02706#pg8>. When you look at Fig. six it shows this relative sMAPE improvement is in a majority of cases, achieving lower sMAPE in sixty-two percent of all comparisons <ref:2603.02706#pg8>.
Taro: So the paper suggests that increasing the number of individual models in ensembles keeps reducing both the average sMAPE and the performance variance, which means it’s getting more reliable <ref:2603.02706#pg8>?. That sounds like a good safety case for real-time systems <ref:2603.02706#pg8>.
Rosa: It really does; the results further indicate that ensemble performance depends primarily on the diversity of the individual models rather than just how complex your combination method is <ref:2603.02706#pg8>. Simple mean or median combinations already yield robust and reliable improvements <ref:2603.02706#pg8>.
Dev: It mitigates that model selection risk, preventing those extreme forecast failures that happen when you pick a single, sub-optimal model <ref:2603.02706#pg8>. The median combination is the most frequent method in the top one hundred ensembles, which makes sense given its robust performance across different models <ref:2603.02706#pg8>.
Taro: So what does this mean for someone who just listens to the show? It means that when you’re looking at long-term power quality trends, relying on one single prediction isn't as safe as using a well-weighted ensemble approach <ref:2603.02706#pg1>.
Rosa: It offers a scalable and computationally efficient framework for transmission system operators to enhance trend identification and support infrastructure planning <ref:2603.02706#pg9>. They are essentially giving operators a tool to better plan their assets using more diverse data <ref:2603.02706#pg1>.
Dev: Before we wrap up, I think the main thing is that ensemble methods effectively reduce error variability by mitigating this model selection risk, which is a big deal for real-time control systems <ref:2603.02706#pg8>. The D28 ensemble achieving that Benchmark Ratio of zero point eight four seven against the best individual model's sMAPE of eighteen point two two percent really shows how much improvement is possible <ref:2603.02706#pg7>.
Taro: I just want to push on this, how does this idea of ensemble forecasting, using diverse models for long-term PQ prediction, actually help someone in the field when they are dealing with real-world system misbehavior <ref:2603.02706#pg1>?.
Rosa: It gives operators a way to spot subtle trends in voltage unbalance or total harmonic distortion that might be missed by one model alone, which is crucial for proactive maintenance planning <ref:2603.02706#pg9>.
Dev: So we’ve seen how the Ensemble Forecasting of Power Quality Parameters works, and it shows that diversity in models leads to more stable and accurate long-term predictions <ref:2603.02706#pg1>.
Taro: I think the key for future work is extending this toward probabilistic forecasting, which would give us a better sense of uncertainty in those predictions <ref:2603.02706#pg9>.
Rosa: Exactly; and we'll keep an eye on that as they look at applying this concept to distribution systems next <ref:2603.02706#pg9>.
Dev: That’s all for this deep dive into the Ensemble Forecasting of Power Quality Parameters, thank you for tuning in <ref:2603.02706#pg1>.
The paper's summary: Rosa: So, to wrap up what we just talked about, this paper is essentially saying that when you're predicting power quality—things like flicker or harmonic distortion—you shouldn't just trust one single prediction model <ref:2603.02706#pg1>.
Dev: Right. It’s confirming that using an ensemble of different models, combining their outputs, is a robust and scalable approach for long-term prediction in transmission systems <ref:2603.02706#pg1>.
Rosa: The core finding is that these ensemble forecasts consistently beat individual models in accuracy and robustness <ref:2603.02706#pg1>. They're seeing significant improvements over those baseline seasonal naive benchmarks, and even over the single best model they tested <ref:2603.02706#pg1>.
Dev: The numbers show that when you look at the top one hundred ensemble models, they actually outperform the absolute best individual model found in that group <ref:2603.02706#pg7>. The D28 ensemble configuration was the overall winner, hitting a low error of about seventeen point six eight percent and a benchmark ratio of zero point eight four seven <ref:2603.02706#pg7>.
Rosa: That benchmark ratio is key. It shows that the ensemble method really stabilizes performance, preventing those crazy forecast failures you often see when you just pick one model that happens to be slightly better for one specific time period <ref:2603.02706#pg8>. It’s about reliability, not just hitting a good score once <ref:2603.02706#pg8>.
Dev: And that stability comes from the way they build the ensemble, using weights based on how accurate each model is, which means the better models get more influence in the final prediction <ref:2603.02706#pg1>. It's a smart way to combine them <ref:2603.02706#pg1>.
Rosa: What this means for us on the ground is that we can use this to improve asset management by enabling proactive maintenance planning and better investment decisions through medium- to long-term PQ prediction <ref:2603.02706#pg9>. We can spot trends before they become big problems <ref:2603.02706#pg9>.
Dev: And the point is, the success of this ensemble approach seems to depend mostly on how diverse the individual models are, not just how complicated your combination method is <ref:2603.02706#pg8>. Simple mean or median combinations already give you reliable improvements <ref:2603.02706#pg8>.
Rosa: Exactly. So, if you’re an operator looking at system stability, this framework gives you a scalable way to enhance trend identification and support infrastructure planning using diverse data sets <ref:2603.02706#pg9>.
Dev: That leads us to thinking about what comes next, because while this paper focuses on accuracy over a fifty-two week horizon, future research will look at extending this toward probabilistic forecasting <ref:2603.02706#pg9>. That would give us a better sense of the uncertainty around those predictions <ref:2603.02706#pg9>.
Rosa: True. They’ll also be looking at applying this idea to distribution systems next, which is a big step for expanding this technology <ref:2603.02706#pg9>. We need to see how well it holds up outside the controlled environment of a big transmission grid <ref:2603.02706#pg1>.
The paper's improvements: Taro: So, we've talked about how ensembles beat single models for accuracy, and now we need to talk about what this research actually suggests we should *do* with that information <ref:2603.02706#pg1>.
Rosa: The main implication is that this ensemble method gives operators a way to spot subtle trends in voltage unbalance or total harmonic distortion that one model just might miss <ref:2603.02706#pg9>. It makes proactive maintenance planning much more feasible <ref:2603.02706#pg9>.
Dev: And it helps with investment decisions because you get a much more stable forecast, which reduces the risk of making a bad guess based on one outlier prediction <ref:2603.02706#pg1>. The D28 ensemble achieving that benchmark ratio of zero point eight four seven really shows how much improvement is possible in terms of reliability <ref:2603.02706#pg7>.
Rosa: It changes the game for long-term planning because it provides a framework where you can use diverse data sources to build a plan that’s less likely to fail when things get weird on the grid <ref:2603.02706#pg9>. It moves us away from just reacting to problems <ref:2603.02706#pg1>.
Dev: Exactly, and the paper suggests that the success of this ensemble method depends primarily on having a good mix of individual models, not on inventing some super complex way to aggregate them <ref:2603.02706#pg8>. Simple mean or median combinations are already yielding robust improvements <ref:2603.02706#pg8>.
Taro: So, the future work they mention is really interesting because they’re planning to look at extending this toward probabilistic forecasting <ref:2603.02706#pg9>. That would give us a better sense of the uncertainty around those predictions <ref:2603.02706#pg9>.
Rosa: Right, and they also want to apply this whole concept to distribution systems next, which is a big step for expanding this technology beyond just the main transmission lines <ref:2603.02706#pg9>. We need to see how well it holds up outside the controlled environment of a big transmission grid <ref:2603.02706#pg1>.
Dev: So, we've seen how ensembles beat single models for accuracy, and now we need to talk about what this research actually suggests we should *do* with that information <ref:2603.02706#pg1>.
Conclusion: Rosa: So to wrap up, this paper on "Ensemble Forecasting of Power Quality Parameters" basically shows that using multiple models together is a way to get much more reliable long-term predictions for power quality <ref:2603.02706#pg1>.
Dev: It confirms that ensemble methods consistently beat individual models in terms of accuracy and robustness, especially when you compare them against those seasonal naive benchmarks <ref:2603.02706#pg1>.
Taro: It’s interesting how they found that the success of this ensemble method depends primarily on the diversity of the individual models, not just on how complicated your combination method is <ref:2603.02706#pg8>. That means we should focus on getting a good mix of different forecasting techniques <ref:2603.02706#pg1>.
Rosa: Exactly, and the results show that ensemble performance depends primarily on the diversity of the individual models rather than on the complexity of the aggregation method <ref:2603.02706#pg8>. It means simple mean or median combinations already yield robust and reliable improvements <ref:2603.02706#pg8>.
Dev: And that means for operators, this approach offers a scalable and computationally efficient framework to help them enhance trend identification and support infrastructure planning <ref:2603.02706#pg9>. It’s a tool for better system management <ref:2603.02706#pg1>.
Taro: I think the biggest thing is that it mitigates that model selection risk by consistently improving forecast accuracy and reducing error variability <ref:2603.02706#pg8>. That stabilization prevents those extreme forecast failures you see when you select just one sub-optimal model <ref:2603.02706#pg8>.
Rosa: It really does; the results further indicate that ensemble performance depends primarily on the diversity of the individual models rather than on the complexity of the aggregation method, as simple mean or median combinations already yield robust and reliable improvements <ref:2603.02706#pg8>.
Dev: So we've seen how ensembles beat single models for accuracy, and now we need to talk about what this research actually suggests we should *do* with that information <ref:2603.02706#pg1>.
Taro: The paper also flags a limitation, though, because they only tested this on transmission system data; the authors themselves flag that the next step is extending this toward probabilistic forecasting <ref:2603.02706#pg9>.
Rosa: Right, and they’ll be looking at applying this whole concept to distribution systems next, which is a big step for expanding this technology beyond just the main transmission lines <ref:2603.02706#pg9>. We need to see how well it holds up outside the controlled environment of a big transmission grid <ref:2603.02706#pg1>.
Dev: So, we've seen how ensembles beat single models for accuracy, and now we need to talk about what this research actually suggests we should *do* with that information <ref:2603.02706#pg1>.
TUD Dresden University of Technology · TransnetBW
eess.SP
Submitted: 2026-03-03
Updated: 2026-03-03
Comments: 5 pages, 7 figures, submitted to IEEE conferences
Journal ref: 2026 22nd International Conference on Harmonics & Quality of Power (ICHQP), Dresden, Germany, 2026, pp. 1-5
DOI: 10.1109/ICHQP67927.2026.11756209
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 80/100
The gist: Ensemble forecasting is confirmed as a robust and scalable approach for long-term PQ prediction in transmission systems The results show that ensemble forecasts consistently outperform individual…
Key concepts
- Ensemble Forecasting
- This method involves creating several individual forecasts from different models and then combining them using a specific function, like a median or weighted average. This approach is used to reduce the risk associated with relying on just one model, leading to more stable and accurate predictions for long-term trends.
- PQ Parameters
- These are measurements of power quality in transmission systems, including flicker severity (Uplt), voltage unbalance (UNB), and total harmonic distortion. Monitoring these parameters helps operators proactively manage assets and detect trends related to electrical environment changes.
- Benchmark Ratio (BR)
- The Benchmark Ratio compares the performance of an ensemble model against the best individual model. A BR less than 1 indicates that the ensemble is superior to the single best model, demonstrating its effectiveness in reducing forecast error and improving overall reliability.
Terminology
Summary
Ensemble forecasting is confirmed as a robust and scalable approach for long-term PQ prediction in transmission systems The results show that ensemble forecasts consistently outperform individual models in terms of accuracy and robustness, achieving significant improvements over seasonal naive benchmarks and the best-performing single models. p. 0 The results show that ensemble forecasts consistently
outperform individual models in terms of accuracy and
robustness, achieving significant improvements over seasonal
naive benchmarks and the best-performing single models. Ensemble forecasting is therefore confirmed as a robust and
scalable approach for long-term PQ prediction in transmission
systems.Where the paper says it
Data Characteristics and Preprocessing
PQ parameters vary over time due to the influence of many different factors, mainly related to the electrical environment Generally, three different types of variations can be distinguished: short-, medium- and long-term variations. The paper focuses on medium- to long-term variations, with long-term referring to time horizons of several months to one year, essential for reliable system performance forecasts. p. 0 The generated energy of PV
systems is significantly higher during the summer months
(longer days, higher angle of insolation). This paper focuses on medium- to long-term variations,
with long-term referring to time horizons of several months to
one year, essential for reliable system performance forecasts.Where the paper says it
The dataset originates from two measurement campaigns within the transmission system of a German and Estonian TSO. p. 0 This paper focuses on medium- to long-term variations,
with long-term referring to time horizons of several months to
one year, essential for reliable system performance forecasts. Dataset
The dataset originates from two measurement campaigns
within the transmission system of a German and Estonian
TSO. Data acquisition employed fixed PQ instruments that comply
with the IEC 61000-4-30 Class A standard. The monitored PQ parameters include different voltage
quality parameters: long-term flicker severity (Uplt), voltage
unbalance (UNB), total harmonic distortion of voltage (Uthd)
and individual harmonic voltages (U03, U05, U07, …, U15)
recorded at a 10-minute aggregation interval.Where the paper says it
Where the paper says it
Data preprocessing involved aggregating the original 10-min measurements to weekly values by computing the 95th percentile for each full calendar week (Monday to Sunday). p. 0 The monitored PQ parameters include different voltage
quality parameters: long-term flicker severity (Uplt), voltage
unbalance (UNB), total harmonic distortion of voltage (Uthd)
and individual harmonic voltages (U03, U05, U07, …, U15)
recorded at a 10-minute aggregation interval. Data Preprocessing
The original 10-min measurements are first aggregated to
weekly values by computing the 95th percentile for each full
calendar week (Monday to Sunday). 5 h of missing data per week in
order to increase data availability. Missing weeks or gaps within the resulting weekly time
series occur frequently, for example due to maintenance To maximize the number of usable time series, a up
to 20 % of missing weeks is allowed. Gaps are filled using the
most recent weekly 95th percentile available within the last 10
weeks. Gaps are filled using the
most recent weekly 95th percentile available within the last 10
weeks. The weekly values are then normalized by relating them to
their respective planning levels, which provides the respective
utilization.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Forecasting Models
This study adopts interpretable models that decompose time series into trend, seasonal, and remainder components because machine and deep learning excel in short-term forecasting but lack long-term efficacy due to reliance on specific structures. The utilized models include Exponential Smoothing (ES), ARIMA/SARIMA, Prophet, and STL decomposition methods. p. 1 Exponential
Smoothing, ARIMA/SARIMA, and Prophet are utilized due
to their computational efficiency, established validation, and
robustness across diverse applications. Overview of Models
One category of forecasting approaches includes simple
baseline methods.Where the paper says it
The eight individual models considered for PQ parameter forecasting are listed in Table I. p. 1 For example, SNaive may be applied
to the seasonal component while ARIMA is used for the
seasonally adjusted series [10]. Selected Models
In total, eight individual models are considered suitable for
PQ parameter forecasting and are listed in Tab. Selected Models
In total, eight individual models are considered suitable for
PQ parameter forecasting and are listed in Tab. OVERVIEW OF INDIVIDUAL FORECASTING MODELS
Model Description
SNaive Seasonal Naïve (benchmark)
HW Holt-Winters Method
SARIMA Seasonal ARIMA
Prophet Additive model with trend and seasonality
STL-Drift STL Decomposition with Drift Method
STL-ES STL Decomposition with Exponential Smoothing
STL-Holt STL Decomposition with Holt’s Method
STL-ARIMA STL Decomposition with ARIMA Method
For the voltage unbalance utilization shown in Fig.Where the paper says it
Where the paper says it
Ensemble Forecasting Framework
Ensemble forecasting is based on the generation of individual (base) forecasts followed by their combination. The paper focuses exclusively on ensembles derived from separate, noninterfering forecasting models. p. 2 Forecasts ensembles can be
generated using approaches such as bootstrapping,
hierarchical forecasting, or independent, non-interfering
models [1]. This paper focuses exclusively on the third category,
considering only ensembles derived from separate, noninterfering forecasting models.Where the paper says it
Ensemble models were constructed from combinations of two models up to all eight, considering all 247 possible configurations. p. 2 I, ensemble models of varying sizes can be constructed,
from combinations of two models up to all eight. All 247
possible configurations are considered, and each ensemble is
assigned a short name in which the letter indicates the
ensemble size (B = 2 models, C = 3 models, …, H = 8 models)
and the number specifies its sequence within that group. Weighted ensembles aim to give greater influence to
better-performing models, for example by assigning weights
based on recent model accuracy or relative performance [1]. In this work, four combination methods are considered and
applied to all 247 ensemble configurations:
Mean (equal-weight averaging of individual forecasts)
Median (pointwise median of individual forecasts)
sMAPE (accuracy-weighted ensemble based on mean
model performance)
Rank (rank-weighted ensemble based on mean model
performance order)
The construction of the not equal weighted ensemble
forecasts 𝑦ොୣ୬ୱ,்ା|் using sMAPE and Rank follows a linear
combination of 𝑀 individual model forecasts 𝑦ො,்ା|்:
𝑦ොୣ୬ୱ,்ା|் = ∑ 𝑤
⋅ ெ
ୀଵ 𝑦ො,்ା|்
To ensure that models with higher accuracy (lower error or
lower rank) receive greater weight, the weights 𝑤
are defined
by the reciprocal of a performance metric 𝜙
normalized
across all models:
𝑤 = 𝜙
ିଵ/ ∑ 𝜙
ெ ିଵ
ୀଵ
The input 𝜙
represents either the mean sMAPE or the mean
rank of model 𝑖 , as determined during the performance
evaluation procedure.Where the paper says it
Where the paper says it
Performance Evaluation
Point-forecast accuracy is evaluated using Mean Absolute Error (MAE) and symmetric Mean Absolute Percentage Error (sMAPE) over a horizon H = 52. For this analysis, performance is classified as good below 10 %, acceptable up to 25 %, and poor or very poor thereafter. p. 2 For a forecast 𝑦ො்ା|் and actual observed value
𝑦்ା over a horizon 𝐻 = 52, sMAPE is defined as:
sMAPE = ଶ
ு
∑
ห௬శି௬ොశ|ห
|௬శ|ାห௬ොశ|ห
ு
ୀଵ
This metric is restricted to 0 % - 200 %, providing a balanced
measure for evaluating forecast performance [12]. For this
analysis, performance is classified as good below 10 %,
acceptable up to 25 %, and poor or very poor thereafter.Where the paper says it
The Benchmark Ratio (BR) is defined for any model i as BRi = sMAPE i / sMAPE baseline. p. 2 These aggregated metrics ( sMAPE തതതതതതതതത and R(sMAPE) തതതതതതതതതതതതതത ) are
subsequently used as the input for the construction of
weighted ensembles as described in Section IV. To assess
the performance relative to the baseline model the Benchmark
Ratio (BR) is defined for any model 𝑖 as:
BR = sMAPE തതതതതതതതത
/sMAPE തതതതതതതതത
ୗୟ୧୴ୣ
Fig. 159
8 HW 5. 199
The results indicate that decomposition-based models
showed clear dominance, with STL-ARIMA and STL-ETS
being the top performers, achieving the lowest mean sMAPE
(18. Notably, the best individual model does not rank within the
top 100 ensemble models, highlighting the robust errorreduction capabilities of forecast combination. The D28
ensemble configuration (consisting of STL-ARIMA, STL-ES,
SARIMA and SNaive) emerges as the overall winner, where
the Median combination achieved the lowest sMAPE of
17. These aggregated metrics ( sMAPE തതതതതതതതത and R(sMAPE) തതതതതതതതതതതതതത ) are
subsequently used as the input for the construction of
weighted ensembles as described in Section IV. To assess
the performance relative to the baseline model the Benchmark
Ratio (BR) is defined for any model 𝑖 as:
BR = sMAPE തതതതതതതതത
/sMAPE തതതതതതതതത
ୗୟ୧୴ୣ
Fig. These aggregated metrics ( sMAPE തതതതതതതതത and R(sMAPE) തതതതതതതതതതതതതത ) are
subsequently used as the input for the construction of
weighted ensembles as described in Section IV. To assess
the performance relative to the baseline model the Benchmark
Ratio (BR) is defined for any model 𝑖 as:
BR = sMAPE തതതതതതതതത
/sMAPE തതതതതതതതത
ୗୟ୧୴ୣ
Fig.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Comparison of Model Performance
The analysis of 988 ensemble models shows significant improvement over individual models. The top 100 models outperforming the best individual model highlights the robust error-reduction capabilities of forecast combination. p. 3 III, ensembles consistently achieve
superior accuracy, with the top 100 models outperforming the
best individual model (STL-ARIMA, sMAPE = 18. Notably, the best individual model does not rank within the
top 100 ensemble models, highlighting the robust errorreduction capabilities of forecast combination.Where the paper says it
The most successful ensembles frequently incorporate the five best individual models into groups of four or five. p. 3 Furthermore, Fig. 5 reveals that the most
successful ensembles frequently incorporate the five best
individual models into groups of four or five. 5 reveals that the most
successful ensembles frequently incorporate the five best
individual models into groups of four or five. While the
Median is the most frequent method in the top 100, the
specific choice of aggregation is ultimately less critical than
the composition and size of the model pool. Ensemble methods effectively mitigate this modelselection risk by consistently improving forecast accuracy and
reducing error variability. The results further indicate that
ensemble performance depends primarily on the diversity of
the individual models rather than on the complexity of the
aggregation method, as simple mean or median combinations
already yield robust and reliable improvements.Where the paper says it
Where the paper says it
Where the paper says it
Conclusion
Ensemble methods effectively mitigate this model-selection risk by consistently improving forecast accuracy and reducing error variability. Consequently, ensemble forecasting offers a scalable and computationally efficient framework for transmission system operators to enhance trend identification and support infrastructure planning. p. 4 The results further indicate that
ensemble performance depends primarily on the diversity of
the individual models rather than on the complexity of the
aggregation method, as simple mean or median combinations
already yield robust and reliable improvements. Consequently, ensemble forecasting offers a scalable and
computationally efficient framework for transmission system
operators to enhance trend identification and support
infrastructure planning. Consequently, ensemble forecasting offers a scalable and
computationally efficient framework for transmission system
operators to enhance trend identification and support
infrastructure planning. Future research will focus on an
application to distribution systems, extending it toward
probabilistic forecasting, and developing asymmetric
accuracy measures that prioritize forecast accuracy near
power quality limits, where prediction errors have the greatest
operational impact. Consequently, ensemble forecasting offers a scalable and
computationally efficient framework for transmission system
operators to enhance trend identification and support
infrastructure planning. Future research will focus on an
application to distribution systems, extending it toward
probabilistic forecasting, and developing asymmetric
accuracy measures that prioritize forecast accuracy near
power quality limits, where prediction errors have the greatest
operational impact.Where the paper says it
Where the paper says it
Where the paper says it
Acknowledgment
This work was supported by the German transmission system operator TransnetBW, the Estonian transmission system operator Elering AS, and the European Union, and was co-funded by the Estonian Ministry of Education and Science (Project No. TemTa-134). p. 4 ACKNOWLEDGMENT
This work was supported by the German transmission
system operator TransnetBW, the Estonian transmission
system operator Elering AS, and the European Union, and was
co-funded by the Estonian Ministry of Education and Science
(Project No. TemTa-134).Where the paper says it
References
[1] X. Wang, R. J. Hyndman, F. p. 4 REFERENCES
[1] X. Hyndman, F. Hyndman, F. Kang, ‘Forecast combinations:
An over 50-year review’, International Journal of Forecasting, vol. Kang, ‘Forecast combinations:
An over 50-year review’, International Journal of Forecasting, vol. 1518–1547, Oct.Where the paper says it
Where the paper says it
Where the paper says it
[2] M. Domagk et al. p. 0 Ensemble Forecasting of Power Quality Parameters
Max Domagk
Peter Feistel
Jan Meyer
TUD Dresden University of
Technology
Dresden, Germany
max. domagk@tu-dresden. Domagk et al. , ‘Trend analysis for power quality parameters based
on long-term measurement campaigns’, in 2022 20th International
Conference on Harmonics & Quality of Power (ICHQP), Naples, Italy,
May 2022, pp.Where the paper says it
Where the paper says it
[3] A. Zeng, M. Chen, L. Zhang, and Q. p. 4 1109/ICHQP53011. Zhang, and Q. Zhang, and Q. Xu, ‘Are Transformers Effective
for Time Series Forecasting? Xu, ‘Are Transformers Effective
for Time Series Forecasting? ’, Proceedings of the AAAI Conference on
Artificial Intelligence, vol. ’, Proceedings of the AAAI Conference on
Artificial Intelligence, vol. 11121–11128, June 2023, doi:
10.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
[4] P. F. Austnes, C. p. 4 11121–11128, June 2023, doi:
10. Austnes, C. Austnes, C. García-Pareja, F. García-Pareja, F. Nobile, and M. Nobile, and M. Paolone,
‘Probabilistic Load Forecasting of distribution power systems based on
empirical copulas’, Sustainable Energy, Grids and Networks, vol.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
[5] C. C. Holt, ‘Forecasting seasonals and trends by exponentially weighted moving averages’, International Journal of Forecasting, vol. p. 4 101708, June 2025, doi: 10. Holt, ‘Forecasting seasonals and trends by exponentially
weighted moving averages’, International Journal of Forecasting, vol. Holt, ‘Forecasting seasonals and trends by exponentially
weighted moving averages’, International Journal of Forecasting, vol. 5–10, 2004, doi: 10.Where the paper says it
Where the paper says it
[6] P. R. Winters, ‘Forecasting sales by exponentially weighted moving averages’, Management Science, vol. p. 4 ijforecast. Winters, ‘Forecasting sales by exponentially weighted moving
averages’, Management Science, vol. Winters, ‘Forecasting sales by exponentially weighted moving
averages’, Management Science, vol. 324–342, 1960.Where the paper says it
Where the paper says it
[7] G. Box, ‘Box and Jenkins: Time Series Analysis, Forecasting and Control’, in A Very British Affair: Six Britons and the Development of Time Series Analysis During the 20th Century, T. p. 4 324–342, 1960. Box, ‘Box and Jenkins: Time Series Analysis, Forecasting and
Control’, in A Very British Affair: Six Britons and the Development of
Time Series Analysis During the 20th Century, T. Box, ‘Box and Jenkins: Time Series Analysis, Forecasting and
Control’, in A Very British Affair: Six Britons and the Development of
Time Series Analysis During the 20th Century, T. ,
London: Palgrave Macmillan UK, 2013, pp.Where the paper says it
Where the paper says it
[8] S. J. Taylor and B. p. 4 1057/9781137291264_6. Taylor and B. Taylor and B. Letham, ‘Forecasting at Scale’, The American
Statistician, vol. Ensemble Forecasting of Power Quality Parameters
Max Domagk
Peter Feistel
Jan Meyer
TUD Dresden University of
Technology
Dresden, Germany
max. domagk@tu-dresden.Where the paper says it
Where the paper says it
Where the paper says it
[9] R. Cleveland and W. p. 4 1080/00031305. Cleveland and W. Cleveland and W. Cleveland, ‘STL: A seasonal-trend
decomposition procedure based on loess’, Journal of Official Statistics,
vol. Cleveland, ‘STL: A seasonal-trend
decomposition procedure based on loess’, Journal of Official Statistics,
vol. 3–73, 1990.Where the paper says it
Where the paper says it
Where the paper says it
[10] K. H. Lem, ‘The STL-ARIMA approach for seasonal time series forecast: A preliminary study’, ITM Web Conf. p. 4 3–73, 1990. Lem, ‘The STL-ARIMA approach for seasonal time series
forecast: A preliminary study’, ITM Web Conf. Lem, ‘The STL-ARIMA approach for seasonal time series
forecast: A preliminary study’, ITM Web Conf. 01008, 2024,
doi: 10.Where the paper says it
Where the paper says it
[11] M. Hibon and T. p. 4 1051/itmconf/20246701008. Hibon and T. Hibon and T. Evgeniou, ‘To combine or not to combine: selecting
among forecasts and their combinations’, International Journal of
Forecasting, vol. Ensemble Forecasting of Power Quality Parameters
Max Domagk
Peter Feistel
Jan Meyer
TUD Dresden University of
Technology
Dresden, Germany
max. domagk@tu-dresden.Where the paper says it
Where the paper says it
Where the paper says it
[12] R. J. Hyndman and G. p. 4 REFERENCES
[1] X. Hyndman, F. Hyndman and G. Athanasopoulos, Forecasting: principles and
practice, 2.Where the paper says it
Where the paper says it
[13] S. Makridakis, E. p. 4 OTexts, 2018. Makridakis, E. Makridakis, E. Spiliotis, and V. Spiliotis, and V. Assimakopoulos, ‘The M4
Competition: 100,000 time series and 61 forecasting methods’,
International Journal of Forecasting, vol.Where the paper says it
Where the paper says it
Where the paper says it
[14] Fig. 3 shows ensemble forecast for the utilization of the voltage unbalance using three different models from Fig. p. 2 To assess
the performance relative to the baseline model the Benchmark
Ratio (BR) is defined for any model 𝑖 as:
BR = sMAPE തതതതതതതതത
/sMAPE തതതതതതതതത
ୗୟ୧୴ୣ
Fig. Ensemble forecast for the utilization of the voltage unbalance
using three different models from Fig. Ensemble forecast for the utilization of the voltage unbalance
using three different models from Fig. 2; training set (105 weeks) in black;
test set in grey (52-week horizon) and ensemble forecast in redWhere the paper says it
Where the paper says it
[15] Fig. 4 shows average sMAPE by ensemble size (top) and model set (bottom) with mean (solid line), 25th-75th percentile range (dark bar) and min-max range (light bar). p. 3 Consequently,
ensemble success appears to be predicated on the inclusion of
diverse, well performing individual models. Average sMAPE by ensemble size (top) and model set
(bottom) with mean (solid line), 25th-75th percentile range (dark bar) and
min-max range (light bar).Where the paper says it
[16] Fig. 5 reveals that the most successful ensembles frequently incorporate the five best individual models into groups of four or five. p. 3 Furthermore, Fig. 5 reveals that the most
successful ensembles frequently incorporate the five best
individual models into groups of four or five.Where the paper says it
[17] Fig. 6 provides illustrative examples of different forecasts for the best individual model and the best ensemble model. p. 4 Comparison of the Best Individual and Ensemble Model
Fig. 6 provides illustrative examples of different forecasts
for the best individual model and the best ensemble model.Where the paper says it
[18] Fig. 7, left shows Empirical cumulative distribution function (ECDF) of relative sMAPE improvements. p. 4 Example forecasts of the best individual model (blue) and best
ensemble (red): training set (105 weeks) in black; test set in grey (52-week
horizon); From top to bottom: the ensemble performs substantially better,
slightly better, and worse than the best individual model
Fig. Comparison of the best individual and the ensemble model:
Empirical cumulative distribution function (ECDF) of relative sMAPE
improvements (left) and scatter plot of individual versus ensemble sMAPE
(right)Where the paper says it
[19] Fig. 7, right shows scatter plot of individual versus ensemble sMAPE. p. 4 Example forecasts of the best individual model (blue) and best
ensemble (red): training set (105 weeks) in black; test set in grey (52-week
horizon); From top to bottom: the ensemble performs substantially better,
slightly better, and worse than the best individual model
Fig. Comparison of the best individual and the ensemble model:
Empirical cumulative distribution function (ECDF) of relative sMAPE
improvements (left) and scatter plot of individual versus ensemble sMAPE
(right)Where the paper says it
Pruning redundant reasoning tokens
The results indicate that decomposition-based models showed clear dominance, with STL-ARIMA and STL-ETS being the top performers. Only these STL variants and SARIMA successfully outperformed the SNaive benchmark (BR < 1), yielding accuracy gains of up to 12. p. 3 67 %) and the best mean ranks. Only these
STL variants and SARIMA successfully outperformed the
SNaive benchmark (BR < 1), yielding accuracy gains of up to
12. Only these
STL variants and SARIMA successfully outperformed the
SNaive benchmark (BR < 1), yielding accuracy gains of up to
12. 7 %, while more complex models like Prophet and HW
failed to surpass the baseline.Where the paper says it
Where the paper says it
How the training signal is built
Consistency is maintained by using a 105-week training period and a 52-week forecast horizon for all time series and models. Accuracy is assessed in three steps: 1. p. 2 Performance Evaluation Procedure
Consistency is maintained by using a 105-week training
period and a 52-week forecast horizon for all time series and
models. Following the approach similar to [13], accuracy is
assessed in three steps:
1. Following the approach similar to [13], accuracy is
assessed in three steps:
1. Calculation of accuracy (MAE and sMAPE) over the
entire horizon for each model and time series. Calculation of accuracy (MAE and sMAPE) over the
entire horizon for each model and time series. Determination of the rank R(sMAPE) for each model per
time series. Determination of the rank R(sMAPE) for each model per
time series. Calculation of the mean accuracy (MAE തതതതതത and sMAPE തതതതതതതതത)
and mean rank (R(sMAPE) തതതതതതതതതതതതതത) for each model across all
716 series. Calculation of the mean accuracy (MAE തതതതതത and sMAPE തതതതതതതതത)
and mean rank (R(sMAPE) തതതതതതതതതതതതതത) for each model across all
716 series. These aggregated metrics ( sMAPE തതതതതതതതത and R(sMAPE) തതതതതതതതതതതതതത ) are
subsequently used as the input for the construction of
weighted ensembles as described in Section IV.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the gains come from
The D28 ensemble configuration emerges as the overall winner, where the Median combination achieved the lowest sMAPE of 17.68 % and a Benchmark Ratio of 0.847. This represents a 2. p. 2 All 247
possible configurations are considered, and each ensemble is
assigned a short name in which the letter indicates the
ensemble size (B = 2 models, C = 3 models, …, H = 8 models)
and the number specifies its sequence within that group. For
example, B01-B28 correspond to ensembles of 2 models,
C01-C56 to ensembles of 3 models, and H01 represents the
ensemble of all 8 models. This represents an
2. 6 % improvement over the best individual model and a
15. 6 % improvement over the best individual model and a
15. 3 % improvement over the SNaive baseline. By
comparing individual versus ensemble sMAPE (Fig. 7, right),
it can be seen that the ensemble stabilizes performance,
preventing the extreme forecast failures often associated with
selecting a single, sub-optimal model.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Ensemble forecasting offers a scalable and computationally efficient framework for transmission system operators to enhance trend identification and support infrastructure planning. Ensemble performance depends primarily on the diversity of the individual models rather than on the complexity of the aggregation method. p. 4 Ensemble methods effectively mitigate this modelselection risk by consistently improving forecast accuracy and
reducing error variability. The results further indicate that
ensemble performance depends primarily on the diversity of
the individual models rather than on the complexity of the
aggregation method, as simple mean or median combinations
already yield robust and reliable improvements. Ensemble methods effectively mitigate this modelselection risk by consistently improving forecast accuracy and
reducing error variability. The results further indicate that
ensemble performance depends primarily on the diversity of
the individual models rather than on the complexity of the
aggregation method, as simple mean or median combinations
already yield robust and reliable improvements.Where the paper says it
Where the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Ensemble forecasting can be integrated into asset management by enabling
proactive maintenance planning and informed investment decisions
through medium- to long-term PQ prediction. -
The ensemble system can provide a robust framework for long-term PQ prediction by mitigating
model selection risk
and reducing error variability, as demonstrated by the D28 ensemble achieving a Benchmark Ratio of 0.847 against the best individual model's sMAPE of 18.22%. -
The improved AI system can perform enhanced trend identification and support infrastructure planning by utilizing ensemble forecasts derived from diverse, noninterfering models, as suggested by the finding that
ensemble success appears to be predicated on the inclusion of diverse, well performing individual models.
Abstract
The growing integration of power electronic-based technologies has increased the necessity of power quality (PQ) monitoring in transmission systems. Although large datasets are collected by operators, their use is typically limited to compliance assessment. Medium- to long-term forecasting can enhance the value of these datasets by enabling proactive asset management and trend detection, despite challenges related to data heterogeneity and seasonality. This paper systematically evaluates individual and ensemble forecasting approaches for PQ parameters in transmission systems. More than 700 weekly time series from measurement campaigns in Germany and Estonia are analysed to assess various models and aggregation strategies within a structured ensemble framework. The results show that ensemble forecasts consistently outperform individual models in terms of accuracy and robustness, achieving significant improvements over seasonal naive benchmarks and the best-performing single models. Ensemble forecasting is therefore confirmed as a robust and scalable approach for long-term PQ prediction in transmission systems.
Related papers
- Open-H-RF: A Large-Scale Ultrasound Raw Channel Dataset Initiative
- Scalable Attention for 5G NR Channel Estimation
- Low-Altitude Wireless Networks: A Comprehensive Survey
- Sensing-Assisted OFDM with Index Modulation: A Waveform for Optical Wireless Integrated Communication and Localisation
- Iteration-Free Candidate Pruning and an Index-Domain VLSI Architecture for Sort-Free MIMO Box Decoding up to 1024-QAM
- A Foundation Model for Large-Scale Wireless Network Planning, Operation and Optimization