Predicting Power Consumption with an Asymmetric Loss and Anomaly Detection Long Short-Term Memory (LSTM) Framework

Abstract

To avoid unwelcome power outages caused by insufficient power output, it is critical to construct a reliable load forecasting model with few underpredictions. It is difficult to anticipate household power consumption trends due to the presence of fluctuations and anomalies in this sector. As a means of increasing the penalty for underpredictions, this study suggests a number of Long Short-Term Memory (LSTM) frameworks that use various asymmetric loss functions. Before the load forecasting work, we additionally use a density-based spatial clustering of applications with noise (DBSCAN) anomaly detection technique to get rid of any existing oultiers. Hourly power consumption, weather, and calendar characteristics are part of the three datasets from France, Germany, and Hungary that are subject to seasonality splitting in order to account for the impact of social and meteorological elements. Relocating the outliers effectively lowers the underestimation and overestimation errors across all seasonal datasets, according to root-mean-square error (RMSE) statistics. Furthermore, seasonality splitting and asymmetric loss functions successfully reduce underestimations while slightly raising the overestimation error. In order to avoid potentially disastrous power outages, it is crucial to decrease underestimations of electrical usage.
Index Terms—Anomaly detection, asymmetric loss, DBSCAN, LSTM, seasonality.