Photovoltaic Power Forecasting Using Neural Networks for Short and Medium-Term Dependencies

Renewable energy sources are pivotal in evolving microgrid systems. Users can generate, utilize, and store electrical power in a local setting by incorporating sustainable energy sources and storage systems. The inherent unpredictability of these sources necessitates innovative technologies like machine learning and the Internet of Things for effective control, upkeep, and integration with the existing electrical grid. A crucial hurdle in this integration, particularly for photovoltaic (PV) systems, is the creation of a precise power forecasting tool. This work introduces a long short-term memory (LSTM) algorithm tailored for multistep-ahead forecasting of PV power. It explores both LSTM and convolutional LSTM (Conv-LSTM) models for short to medium-term PV power prediction. The aim is to enhance PV power forecasting accuracy, extending the time horizon of forecasts while maintaining reasonable error margins and processing efficiency. Various criteria are assessed to demonstrate the effectiveness of the proposed algorithms. Factors like processing duration and machine resource requirements for training and testing phases are key in choosing optimal neural network parameters. Results from this study highlight the efficacy of the LSTM model in short and medium-term PV power forecasting, showcasing its potential in renewable energy integration into the power grid.