Short Term Residential Load Forecasting Using Temporal Weather Based Embedding Stacked LSTMs

Abstract

Resource management is crucial to balance human needs with sustainability, prevent overuse, and preserve natural resources like water, forests, and minerals for future generations. Managing electricity at the root of human usage can be a crucial first step, helping us move toward better resource management and reducing the strain on natural resources. Superior forecasting approaches are needed to determine usage patterns. Accurate predictions can serve as key input to the Home Energy Management Systems (HEMS) mechanisms in optimizing electricity operation, reducing energy waste, and increasing resource utilization. Neural network-based methods are being developed to forecast electricity usage in residential buildings by learning behavioral patterns over time. These approaches leverage historical data to identify trends and predict future consumption, offering a promising direction for more accurate forecasting methods. Although still evolving, they provide a foundation for optimizing energy management by anticipating demand and enabling more efficient resource allocation. However, these approaches primarily rely on historical patterns to predict future electricity usage, often overlooking the impact of daily weather conditions. In this paper, we explore a method that incorporates weather information to enhance electricity usage predictions. We propose a simple Stacked LSTM-based neural network that integrates historical usage data and weather information as learned inputs for more effective electricity usage prediction. Our approach demonstrates improved prediction performance compared to methods that do not account for weather factors and the CNN-SLSTM model. For the BR04 hourly test dataset, our proposed model achieves a 56% and 67% reduction in RMSE compared to the SLSTM with weather and CNN-SLSTM models, respectively.