Rolling and Day-Ahead Forecasting of Electricity Market Prices: Evaluating the Performance of DL Models with Bayesian Optimization

This paper investigates the application of deep learning (DL) algorithms for short-term electricity market price forecasting which is crucial for cost-effective operation of power systems. We propose four DL models that combine different architectures, including convolutional neural networks (CNNs) and long short-term memory (LSTM) networks. This further involves a Bayesian Optimization (BO) method for hyperparameter tuning of the two best models to ensure optimal performance, as these models are sensitive to hyperparameters. Moreover, to evaluate the performance of the model, this paper also presents rolling forecasting using the LSTM model. Forecasting capability of the DL models is evaluated by using accuracy measures such as mean absolute percentage error (MAPE), root mean square error (RMSE), and mean absolute error (MAE). Further, the DL models are also compared with our previous work. This paper utilizes the hourly data of prices and loads acquired from the PJM electricity market. Our findings indicate that CNN+LSTM and CNN+BiLSTM models outperform the other models for day-ahead forecasting, whereas BiLSTM and CNN+BiLSTM models show better performance for week-ahead price forecasting.