Short-Term and Rolling Solar PV Power Forecasts: Performance Evaluation and Hyperparameter Tuning of Deep Learning Models

Accurate and dependable forecasting of PV power is of utmost importance in commercial PV plants. Nevertheless, the unpredictable and intermittent nature of solar radiation poses challenges for accurate predictions. Inspired by this, we present hybrid deep learning models, which utilize Long Short-Term Memory (LSTM), Bidirectional LSTM (BiLSTM), and Temporal Convolution Networks (TCN s), to generate short-term solar PV power forecasts for multiple season and weather conditions. The models' initial forecasts were first assessed using accuracy measures to determine the best performing model(s) followed by their hyperparameter tuning using Bayesian optimization (BO) and Random Search optimization (RSO). After successful optimization, additional case studies were conducted to demonstrate the improved forecasting accuracy of the optimized models. It is shown that the performance of BO, LSTM-TCN and RSO, Bi-LSTM-TCN are more suitable for 72hr-ahead forecast. Additionally, an intra-day rolling forecast was performed utilizing one of the base models (LSTM) to demonstrate its potential for achieving high accuracy in PV power forecasts. The obtained results provide valuable insights into the effectiveness of the proposed models and their application in real-world scenarios.