Data-Driven Load Prediction Models for Distribution Systems and Transformer Station Areas

The forecasting accuracy of electrical load in distribution systems and transformer stations is essential to increase stability and reliability of the grid. Therefore, this research paper proposes a new hybrid forecast model based on temporal deep learning and feature refining techniques to forecast electrical load in distribution systems for a short-term period. In this case, data on historical electrical demand, meteorological variables and power network topology information are obtained from PJM Hourly Energy Consumption Dataset, MATPOWER IEEE Bus Dataset and NOAA Climate Dataset, correspondingly. For the purpose of improving data quality and stability before the learning process, preliminary preprocessing is conducted through interpolation, moving average, Savitzky-Golay filters, Zscore outlier detection, hourly time synchronization and Min-Max normalization. Afterwards, operational characteristics which consist of temporal, weather and electric components are extracted. Selection of features is conducted by means of Correlation Analysis (CA). Redundant variables are identified using Principal Component Analysis (PCA). Further, Sliding Window (SW) technique is employed to construct temporal sequence. In the forecasting architecture, there are two components: hybrid LSTM-XGBoost model, Long Short-Term Memory (LSTM) model used for learning temporal dependency and XGBoost regression algorithm for nonlinear load prediction. Also, hyperparameter tuning through Harmony Search Algorithm (HSA) is applied for improving the convergence rate and prediction accuracy. From experimental validation, it is evident that the proposed model outperforms in terms of accuracy and efficiency in predicting the model occurrence as compared to RF and SVM models with a MAE of 0.0210, MSE of 0.0006, MAPE of 0.0507, and R² of 0.9623.