A hybrid framework for day-ahead electricity spot-price forecasting: A case study in China

The electricity price volatility can be aggravated by multiple factors, such as load pattern, line limit, regulations, renewable energy generations, weather conditions and holiday. Due to these complex dynamic characteristics of electricity prices, highly accurate forecasting is quite challenging. Our objective is to provide a hybrid framework to forecast 96-point day-ahead electricity price for the following day. We first conducted a day similarity algorithm (DSA) to construct features from electricity price corresponding to the similar days. A deep neural network (DNN) model was developed from 60 important features, including supply, demand and similar day characteristics, selected by eXtreme Gradient Boosting (XGBoost) algorithm. The hyperparameters were tuned using adaptive TPE (ATPE) method. The framework was validated in the real-world dataset of electricity spot market in Shandong Province in China. The proposed framework had good forecasting performance with the lowest MAE of 0.138, MSE of 0.028, RMSE of 0.166 and U2 of 0.434 in the test set and outperformed the other models significantly. The framework presented a robust methodology for market participants to forecast electricity prices accurately, increase profits and improve decision-making skills.