Day ahead load forecasting for campus buildings with solar PV using a hybrid convolutional neural network and bidirectional long short-term memory model

Educational buildings are significant contributors to national energy consumption, with load patterns that fluctuate significantly between semester and non-semester periods, posing challenges for efficient energy management. However, this unique operational structure also presents substantial opportunities for cost savings and income generation. For instance, during semester breaks when energy demand is lower, surplus electricity generated from on-site solar photovoltaic (PV) systems can be exported to the grid, offering potential financial returns. This highlights the importance of accurate load forecasting to support informed energy scheduling and maximize economic benefits. Furthermore, the integration of solar PV introduces additional complexity to energy management due to the inherent variability and intermittency of solar generation. These characteristics require more precise forecasting methods to ensure efficient self-consumption, minimize reliance on the grid, and minimize cost. Therefore, accurate day-ahead load forecasting is essential not only for improving operational efficiency but also for enabling educational institutions to fully capitalize on the economic and sustainability benefits of solar energy. Existing forecasting models often rely on consistent data patterns or necessitate separate models for different periods, which can be inefficient. While various forecasting methods have been explored, including Long Short-Term Memory (LSTM) and Bidirectional LSTM (BiLSTM), research on their application in campus settings remains limited and mostly separating the model into different semester status which is inconvenient for real time application. This study proposes a novel load forecasting approach tailored for both non-semester and semester periods, combining Convolutional Neural Network (CNN) with BiLSTM to improve feature extraction. Three distinct models were developed for validation, namely, Artificial Neural Network (ANN), BiLSTM, and hybrid CNN with BiLSTM (CNN-BiLSTM). The key input features include load consumption, hour of the day, calendar factors (such as holidays), and previous day and week lagged load data. The models were developed using datasets collected at 30-minute intervals over six months from three Universiti Teknologi MARA (UiTM) campuses namely UiTM Permatang Pauh, UiTM Dungun, and UiTM Alor Gajah. Their performance was evaluated using Mean Absolute Percentage Error (MAPE), Root Mean Square Error (RMSE), Mean Absolute Error (MAE), Mean Square Error (MSE) and Coefficient of Determination (R²) on both semester run and combined semester period datasets. The performance evaluation of ANN, BiLSTM, and the proposed CNN-BiLSTM model across UiTM Permatang Pauh, Alor Gajah, and Dungun datasets demonstrates that CNN-BiLSTM consistently outperforms the other models. In the combined dataset, CNN-BiLSTM achieved the highest R² values of 0.935, 0.924, and 0.926, surpassing BiLSTM and ANN. Additionally, it achieved the lowest MAPE, reducing errors by up to 26.66% compared to ANN and 5.10% compared to BiLSTM. The combined dataset approach further enhanced model generalization by capturing seasonal variations and transition phases, making CNN-BiLSTM a more robust solution for load forecasting in university campuses.