LSTM with temporal encoding for irregular time series forecasting in power consumption

Abstract

Power consumption data obtained from sensors are often recorded at irregular time intervals due to network disruptions, device errors, or power outages, resulting in irregular time series that make forecasting difficult. This study aims to develop an electricity consumption forecasting model based on Long Short-Term Memory (LSTM) and Temporal Encoding.  LSTM was chosen because it has an effective gating mechanism for capturing temporal dependencies in time series data, while Temporal Encoding explicitly represents time information to handle irregular time intervals without data imputation.  The methods in this study include data collection via four electrical current sensors, followed by data aggregation every 10 minutes, and feature engineering using sinusoidal encoding and a time difference encoder. The features were normalized using min-max scaling, organized into a multivariate sequence using a sliding window, and divided using a holdout scheme. The model was trained using LSTM and evaluated using Mean Squared Error (MSE). The results show training MSE values of 9.89210^-4, 7.34910^-4, 9.53510^-4 and 1.90610^-3, while the testing MSE values are 4.56610^-3, 2.99310^-3, 1.09410^-2 and 1.20910^-2 for each node. These findings indicate that temporal encoding performs well on the training data, but the model's generalization ability remains limited.