A framework for multi-stage ML-based electricity demand forecasting

This paper proposes a novel framework for energy utility companies to anticipate their customers' energy usage based on their historical consumption data. The proposed framework comprises three major stages: (i) it detects and removes anomalies in consumers' energy consumption data by employing the isolation forest (iForest); (ii) it forms clusters of distinct consumer groups based on similarities in their consumption behavior via the k-means clustering algorithm; and (iii) it predicts electricity consumption by using deep learning algorithms. To this end, two different deep learning algorithms are designed: a long short-term memory (LSTM) network and the combination of convolutional neural network (CNN) and LSTM (referred to as CNN-LSTM) with multiple inputs. Since the latter is a combination of CNN and LSTM models, we apply a 2-D discrete wavelet transform (DWT) based feature extraction to the Gramian angular field (GAF) transformation of the time series to improve the accuracy of predictions. Various evaluation metrics are utilized for 1-hour- and 24-hours-ahead predictions with two different sliding-window sizes, i.e., 24 hours and 36 hours. The results demonstrate that the CNN-LSTM performs significantly better in predicting 24-hours-ahead electricity consumption.