Solar Power Forecast Using Artificial Neural Network Techniques

Solar energy is a clean, available and renewable source of energy. Deployment of photovoltaic panels is trending to take advantage of solar energy for electrical power generation. Intermittent nature of solar energy results in variable generated power. This uncertainty could be alleviated by taking advantage of energy storage systems and accurate solar power forecast. This thesis aims to implement a solar forecast module to take part in an optimized Energy Management System (EMS). Different solar power prediction methods are studied including statistical methods, sky imagers, satellite imaging and Numerical Weather Prediction (NWP). Artificial neural networks (ANNs) which are a subset of statistical methods are chosen as prediction method to satisfy requirements imposed by EMS. The requirements include precise forecast in short-term prediction horizon for proper functionality of the EMS. Prediction horizon is amount of time in future for which prediction is needed. Multiple ANN architectures suitable for time-series prediction are investigated and compared, including Long Short-term Memory (LSTM), Long- and Short-term Time Series Network (LSTNet) and Temporal Convolutional Network (TCN). Although all the aforementioned methods have acceptable performance, TCN architecture shows more promising results. In order to further improve the prediction accuracy, effect of clustering the dataset into sunny and cloudy sub-datasets and using a dedicated prediction module for each subdataset is studied. Results show that prediction accuracy is improved by clustering the dataset for all the models. Moreover, simulation results on datasets of multiple geographical locations with different climate conditions show that prediction accuracy is higher in locations having more stable weather and sunny days. A Graphical User Interface (GUI) is implemented to simplify working with the forecast module. Service is made available to users through exposing REST APIs on a remote server to facilitate user interactions with the forecast service.