An Approach to Improve the Performance of PM Forecasters

摘要: The particulate matter (PM) concentration has been one of the most relevant environmental concerns in recent decades due to its prejudicial effects on living beings and earth’s atmosphere. High PM affects human health several ways leading short long term diseases. Thus, forecasting systems have developed support decisions organizations governments alert population. Forecasting based Artificial Neural Networks (ANNs) highlighted literature their performances. In general, three ANN-based approaches found for this task: ANN trained via learning algorithms, hybrid that combine search algorithms with ANNs, other forecasters. Independent approach, it is common suppose residuals (error series), obtained from difference between actual series forecasting, a white noise behavior. However, possible assumption infringed to: misspecification model, complexity time or temporal patterns phenomenon not captured by forecaster. This paper proposes an approach improve performance forecasters modeling. analyzes remaining recursively patterns. At each iteration, if there are residuals, generates order series. proposed can be used either only forecaster combining two more models. study, system (HS) composed genetic algorithm (GA) modeling performed methods, namely, own system. Experiments were PM2.5 PM10 Kallio Vallila stations Helsinki evaluated six metrics. Experimental results show improves accuracy method terms fitness function all cases, when compared without correction. correction HS superior performance, reaching best five out These also sensitivity analysis was varying proportions sets training, validation test. reached consistent correction, showing interesting tool