Incorporating spatio-temporal population dynamics into deep learning models for small-area population forecasting

Small-area population forecasting has become increasingly critical for establishing localized policies to address regional demographic disparities. Due to its minimal data requirements and methodological simplicity, trend extrapolation offers a practical approach given the limited availability of detailed demographic data. The Long Short-Term Memory (LSTM) model enhances this approach by effectively capturing complex non-linear temporal dependencies. However, forecasting accuracy is often limited by ignoring spatial interdependence, as demographic shifts are frequently driven by interactions between neighboring areas rather than occurring independently within each area. This study evaluates the impact of integrating spatio-temporal neighboring population dynamics into a deep learning model for small-area population forecasting. Specifically, it compares a proposed Convolutional Neural Network (CNN)-LSTM model, which integrates interregional population changes by employing one-dimensional convolutional layer as spatial feature extractor for neighboring dynamics, against the standard LSTM model that treats each unit as an isolated series. The results demonstrate that the CNN-LSTM model generally outperforms the standard LSTM model in overall global accuracy, primarily driven by its performance in areas with high spatial association. This advantage is particularly evident in areas with low temporal but high spatial association, where information from neighbors compensates for discontinuous historical trends. In contrast, caution is required when applying the CNN-LSTM in areas characterized by high temporal and low spatial association, as uncorrelated spatial information from neighbors could act as noise over long-term forecasting.