An Approach to Improve the Performance of PM Forecasters.

Area of Science:

• Environmental Science
• Data Science
• Atmospheric Science

Background:

• Particulate matter (PM) concentration poses significant environmental and health risks.
• Accurate PM forecasting is crucial for public health alerts and environmental management.
• Existing Artificial Neural Network (ANN) forecasting models often assume residual errors follow white noise, which may not hold true.

Purpose of the Study:

• To propose and evaluate a novel approach for improving PM forecasting performance by modeling residual errors.
• To investigate the effectiveness of analyzing temporal patterns in residuals to correct forecasting inaccuracies.
• To assess the proposed method's impact on hybrid systems combining genetic algorithms (GA) and ANNs for PM forecasting.

Main Methods:

• Developed a recursive residuals modeling approach to identify and correct temporal patterns in forecast errors.
• Applied the residual modeling technique to a hybrid system (HS) integrating GA and ANN for PM2.5 and PM10 forecasting.
• Evaluated the enhanced forecasting performance using six metrics across different data splits for training, validation, and testing.

Main Results:

• The proposed residual modeling approach consistently improved the accuracy of PM forecasting compared to uncorrected methods.
• The hybrid system-based residual correction demonstrated superior performance, achieving the best results in the fitness function and five out of six evaluation metrics.
• Sensitivity analysis confirmed the robustness of the approach across varying data set proportions.

Conclusions:

• Residual modeling is an effective strategy for enhancing the accuracy of particulate matter forecasting systems.
• The proposed recursive approach offers a valuable tool for refining predictions from single or combined forecasting models.
• The study highlights the potential of advanced error analysis in environmental forecasting applications.