SAC-ConvLSTM: A novel spatio-temporal deep learning-based approach for a short term power load forecasting

The short-term spatiotemporal load forecasting of power distribution networks has been extensively studied in light of high importance in distribution system control and consumption management. Load forecasting is dependent on temporal and spatial parameters e.g., temperature, rainfall, and land use. Additionally, load time series have spatial autocorrelation in cities. Deep learning has been recently demonstrated to be effective and efficient in time-series and load forecasting in power distribution networks. The present study proposes a new algorithm based on Spatial Auto Correlation and Convolutional Long Short-Term Memory (SAC-ConvLSTM) algorithm via spatiotemporal power load autocorrelation modeling and deep learning. Load time-series signals in a power distribution network are decomposed into sub-signals using the discrete wavelet transform (DWT). Then, the spatio-temporal autocorrelation of sub-signals is calculated, alleviating prediction model complexity using spatial statistics. The areas with significant positive/negative/ no significant autocorrelation in each time-series sub-signal are separately used as the input of the SAC-ConvLSTM algorithm. Finally, the sub-signals are reconstructed into the load signals. This algorithm models the spatial relationships in time-series at different time intervals for power consumption forecasting. The proposed methodology was found to outperform the support vector machine (SVM), Long Short Term Memory (LSTM), Convolutional LSTM (ConvLSTM), Convolutional Gated Recurrent Unit (ConvGRU), Planar Flow-Based Variational Auto-Encoder (PFVAE) and Federated Averaging (FedAVG) algorithms in short-term power load forecasting for short term forecasting of up to two weeks, based on the evaluation metrics of root-mean-square error (RMSE) and mean absolute error (MAE), producing values of 11.08% and 7.02% respectively.