Interpretable PM2.5 Forecasting for Urban Air Quality: A Comparative Study of Adaptive Time-Series Models

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Hi, I'm Tom.

Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.

Tom: We're looking at the paper: "Interpretable PM2.5 Forecasting for Urban Air Quality".

Jane: Here's the gist of the paper. The gist The study investigates whether lightweight and interpretable forecasting approaches can provide competitive performance for hourly PM2.5 prediction in Beijing, China.

Tom: The researchers formulated hourly PM2.5 forecasting for Beijing as a multistep time-series prediction task with exogenous drivers <ref:2603.25495#pg11> Given observations up to time t, the objective is to forecast the next 7 days (168 hours) <ref:2603.25495#pg11> They conducted experiments under the Perfect Prognosis (PP) assumption, where future values of the exogenous regressors over the forecast horizon are treated as known <ref:2603.25495#pg11> The data were kept in chronological order and divided into 90% training and 10% test segments <ref:2603.25495#pg11> Performance is evaluated on the test period using a rolling weekly protocol (horizon = 7 days, step = 7 days), producing multiple weekly forecast windows whose errors are aggregated <ref:2603.25495#pg11> Data Preprocessing and Feature Selection The data preprocessing involved several steps to ensure leakage-safe normalization and robust series handling <ref:2603.25495#pg11> All data streams were temporally aligned and sorted by timestamp, and a continuous hourly index was enforced to ensure a regular sampling grid <ref:2603.25495#pg11> To reduce the influence of extreme values while preserving sample size, outliers were handled via winsorization using training-set percentile bounds <ref:2603.25495#pg11> For leakage-safe normalization, exogenous regressors were standardized using statistics computed only from the training segment and the same transformation was applied to the test segment <ref:2603.25495#pg11> Feature selection was performed using three complementary filter-based criteria: Pearson correlation, mutual information (MI), and minimum redundancy–maximum relevance (mRMR) <ref:2603.25495#pg11> The analysis showed that gaseous precursors were consistently among the most relevant predictors, whereas meteorological variables exhibited comparatively weaker influence <ref:2603.25495#pg11> Based on these considerations, the final exogenous feature set used for PM2.5 forecasting was identified as Forecasting Models Compared Three forecasting model families were benchmarked: SARIMAX, Facebook Prophet (FBP), and NeuralProphet (NP) <ref:2603.25495#pg11> SARIMAX models the target as a linear function of its past values, seasonal dynamics, and contemporaneous external covariates <ref:2603.25495#pg11> Facebook Prophet is a decomposable additive model that represents the series as the sum of interpretable components, typically including a piecewise trend, multiple seasonalities, and the effect of external regressors <ref:2603.25495#pg11> NeuralProphet extends the Prophet formulation by learning decomposable components using neural network parameterizations and by optionally incorporating autoregressive structure to better capture short-term dynamics <ref:2603.25495#pg11> Adaptive Forecasting Regimes The study evaluated each forecasting model under two adaptive deployment regimes <ref:2603.25495#pg11> Regime 1 is the weekly walk-forward refit, where the model is re-estimated at the start of every test week using all data observed up to that point <ref:2603.25495#pg11> Regime 2 is a frozen base model with online residual correction, where adaptation is achieved through a lightweight online correction applied to the model outputs using recently observed forecast errors <ref:2603.25495#pg11> For the frozen-model setting, the corrected forecast for week w is formed as yˆfinal t = ˆy base t + bw, t ∈ week w <ref:2603.25495#pg11> Key Findings on Performance Under weekly walk-forward refitting, Facebook Prophet (FBP) achieved the strongest overall performance, with an MAE of 37.61 and an RMSE of 50.10 <ref:2603.25495#pg11> FBP was also markedly more efficient, completing in 15 min 21.91 s compared to NeuralProphet (NP) which required 88 min 35.19 s <ref:2603.25495#pg11> In the frozen-model setting, SARIMAX showed the largest gain, with MAE decreasing from 47.14 to 32.50 and RMSE decreasing from 65.45 to 46.85 <ref:2603.25495#pg11> Notably, corrected Facebook Prophet reached nearly the same overall error as its walk-forward counterpart while reducing execution time from 15 min 21.91 s to only 46.60 s <ref:2603.25495#pg11> Overall, the study shows that lightweight additive models can provide strong predictive performance with substantially lower operational cost <ref:2603.25495#pg11> The findings show that transparent and computationally efficient forecasting strategies can provide robust air-quality predictions, making them well-suited for real-world deployment where interpretability, low overhead, and ease of maintenance are important considerations.

Jane: Limitations Several limitations should be acknowledged <ref:2603.25495#pg11> First, the dataset was obtained through public API services that provide aggregated pollutant and meteorological estimates <ref:2603.25495#pg11> Second, the reported results reflect Beijing’s emission profile, seasonal pollution cycles, and local meteorological dynamics and may not transfer directly to cities with different source compositions, atmospheric conditions, or urban structures <ref:2603.25495#pg11> Finally, although Prophet-based approaches are generally robust to recurring temporal patterns and gradual drift, substantial discontinuities or abrupt regime shifts could still affect their stability and forecasting reliability <ref:2603.25495#pg11> This study addressed the problem of short-term urban air-quality forecasting by examining whether lightweight and interpretable forecasting frameworks can remain competitive against more complex and computationally demanding approaches <ref:2603.25495#pg11> The findings show that lightweight additive modeling remains a strong and practical alternative for operational air-quality forecasting <ref:2603.25495#pg11> Facebook Prophet achieved the most favorable balance between predictive accuracy, stability, and computational efficiency across both regimes <ref:2603.25495#pg11> Overall, the study shows that transparent and computationally efficient forecasting strategies can provide robust air-quality predictions, making them well-suited for real-world deployment where interpretability, low overhead, and ease of maintenance are important considerations.

Tom: Data Availability All data and code required to reproduce the findings of this study are publicly available in the project GitHub repository <ref:2603.25495#pg11> Acknowledgments The authors have no acknowledgments to declare <ref:2603.25495#pg11> References The references listed support the claims made throughout the paper, such as those regarding air pollution forecasting and specific model comparisons <ref:2603.25495#pg11>.

Title and authors: Tom: So, what are we actually comparing? The paper evaluates three forecasting families: SARIMAX, Facebook Prophet, and NeuralProphet <ref:2603.25495#pg1>. Jane That’s the classic statistical model with SARIMAX, the decomposable additive model which is Facebook Prophet, and then NeuralProphet which extends Prophet using neural network parameterizations for better short-term dynamics <ref:2603.25495#pg1>.

Lu: The tension they highlight in their related work is between predictive performance and the transparency required for policy use <ref:2603.25495#pg3>, and this study directly addresses that by comparing these different model families across two deployment regimes: a weekly walk-forward refit and a frozen base model with online residual correction <ref:2603.25495#pg1>.

Meng: That adaptive part sounds crucial for real systems. If the system drifts over time, does it just break or can it adjust? Jane They are testing both ways, so Regime one is where the model gets re-estimated at the start of every test week using all data up to that point <ref:2603.25495#pg1>, and Regime two is a frozen base model with an online correction applied using recently observed forecast errors <ref:2603.25495#pg1>.

Tom: They also look at feature selection too, showing that gaseous precursors were consistently among the most relevant predictors, whereas meteorological variables exhibited comparatively weaker influence <ref:2603.25495#pg1>. Jane That’s important because it tells us which inputs actually matter for the PM2 point 5 prediction in Beijing <ref:2603.25495#pg1,PM2.5 prediction in Beijing>.

Lu: The paper explores this evolution from classical statistical methods to deep learning and hybrid frameworks, noting that while deep learning gets richer temporal dependencies, there’s a persistent tension between predictive performance and the transparency required for policy use <ref:2603.25495#pg3>.

The paper's summary: Tom: The authors suggest some specific ways to improve the existing approaches by focusing on how adaptive strategies can maintain performance under drift without needing constant, full retraining <ref:2603.25495#pg1>. Jane They introduce two distinct regimes for deployment, and the idea is that this allows systems to adapt to level shifts and gradual drift through an online bias term in the frozen setting <ref:2603.25495#pg1>.

Lu: And they also found that using corrected SARIMAX in the frozen regime showed a significant gain, decreasing its MAE from forty-seven point one four down to thirty-two point five zero and RMSE from sixty-five point four five to forty-six point eight five <ref:2603.25495#pg1>.

Meng: That’s a big win for stability if you’re deploying something that needs to run continuously without constant heavy retraining cycles, but what about the speed? Tom That's where things get interesting because they also looked at execution time, and they found that corrected Facebook Prophet reached nearly the same overall error as its walk-forward counterpart while reducing its execution time from fifteen minutes and twenty-one point nine one seconds down to only forty-six point six zero seconds <ref:2603.25495#pg1>.

Jane: So, it’s not just about accuracy; they are showing that these lightweight additive models can offer a strong predictive performance with substantially lower operational cost, which is a huge deal for real-world deployment <ref:2603.25495#pg1>.

Lu: This suggests that simple adaptive mechanisms can substantially improve performance without repeated full-model retraining when you look at the frozen model setting.

The paper's improvements: Tom: To wrap up the "Interpretable PM2 point 5 Forecasting for Urban Air Quality: A Comparative Study of Adaptive Time-Series Models," the paper shows that lightweight additive modeling remains a strong and practical alternative for operational air quality forecasting <ref:2603.25495#pg1>. Jane Specifically, Facebook Prophet achieved what they called the most favorable balance between predictive accuracy, stability, and computational efficiency across both the walk-forward and frozen regimes <ref:2603.25495#pg1>.

Lu: I think this result supports the idea that transparent and computationally efficient forecasting strategies can provide robust air quality predictions when you need interpretability alongside low overhead <ref:2603.25495#pg1>.

Meng: From my side, the finding that Prophet's correction cuts execution time by over fifty percent is what matters for production systems; it makes it much easier to scale without increasing infrastructure costs.

Lalam: I’m confident in this direction. The study shows that these transparent and computationally efficient strategies give us robust air quality predictions, making them well-suited for real-world deployment where interpretability and low overhead are important considerations <ref:2603.25495#pg1>.

Tom: Well, that's what we had there with the "Interpretable PM2 point 5 Forecasting for Urban Air Quality: A Comparative Study of Adaptive Time-Series Models <ref:2603.25495#pg1,Interpretable PM2.5 Forecasting for Urban Air Quality: A Comparative Study of>." It shows that when you need to predict something critical like air quality, you don't always have to use the heaviest AI models available; a well-tuned additive model can be very effective and much cheaper to run than some of those deep learning alternatives.

Jane: Exactly. The whole point is that we get a transparent result that's accurate enough for policy planning while keeping the computational load down, which makes it much easier for city planners to actually use the forecast in their daily work without needing massive server resources.

Lu: It really highlights that the choice of model isn't just about chasing the absolute highest accuracy number; it’s about finding that sweet spot where you have both a reliable prediction and something you can actually explain to someone who isn't a data scientist.

Meng: And from my perspective, when we ship models, operational cost and speed are huge constraints; if this method cuts the processing time by over half compared to other methods, that translates directly into less infrastructure needed for real-time monitoring across a whole city.

Lalam: I’m confident in this direction. The study shows that these transparent and computationally efficient strategies give us robust air quality predictions, making them well-suited for real-world deployment where interpretability and low overhead are important considerations <ref:2603.25495#pg1>.

Conclusion: ---: Conclusion ---

Tom: So we've looked at "Interpretable PM2 point five Forecasting for Urban Air Quality: A Comparative Study of Adaptive Time-Series Models," and the main thing is that lightweight additive modeling is a strong alternative for operational air quality forecasting one <ref:2603.25495#pg2>. Jane Exactly, and Facebook Prophet seemed to find the best balance between accuracy, stability, and how fast it runs in both the retraining setup and the frozen setup one.

Lu: I think what this study really shows is that transparent models with low overhead can give us robust air quality predictions when we need interpretability alongside minimal computational cost one.

Meng: That speed thing is key. If Prophet's correction cuts execution time by over fifty percent, that means you can scale this without needing massive infrastructure for real-time monitoring across a whole city.

Lalam: I’m confident in this direction. The study shows that these transparent and computationally efficient strategies give us robust air quality predictions, making them well-suited for real-world deployment where interpretability and low overhead are important considerations one.

Tom: Yeah, so when we look at the whole thing, even with all those adaptive regimes like walk-forward refit versus frozen correction, Prophet holds its own better than some of the more complex options we compared one.

Jane: It really shows that you don't always need to go straight to the most computationally expensive deep learning framework if a simpler, interpretable one can handle the task reliably.

Lu: It confirms that simple adaptive mechanisms can substantially improve performance without repeated full-model retraining when you look at the frozen model setting one.

Tom: Well, it's good to see how these time-series models are being adapted for real environmental problems like PM2 point five prediction in Beijing <ref:2603.25495#pg2>.

Jane: It sets a really clear benchmark for what makes a forecasting system truly useful in a public health context.

Lu: Next time, we might look at how this kind of interpretability translates into other complex areas, like how models adapt to domain shifts when you move from one dataset to another.