Industrial and urban growth in Aguascalientes, Mexico, has led to elevated PM \(_{2.5}\) levels, posing significant public health risks. Given limitations in local monitoring infrastructure, this study develops a feedforward neural network model to predict PM \(_{2.5}\) concentrations, incorporating temporal, climatic, and spatial data, as well as prior PM \(_{2.5}\) levels. To address spatial data gaps, we apply inverse distance weighted (IDW) interpolation, enhancing data representation in areas far from monitoring stations. To ensure reliable predictions when previous system states are unknown, we propose alternative estimation methods, including spatial averaging and model training independent of prior states. Model validation uses k-fold cross-validation with metrics like the coefficient of determination, mean squared error, and mean absolute error. Results show that integrating IDW and selectively adjusting temporal variables increases model stability and accuracy. This adaptable forecasting model offers practical applications for regions with limited monitoring resources and sets the stage for future studies incorporating more complex geographic and environmental data.
