Neural Network Based Prediction of Solar Power Generation

Abstract

This project shows the development of solar power forecasting by using a recurrent neural network-based time-series model, typically a Nonlinear Autoregressive model with External Inputs (NARX) to predict the solar power produced by using real-time environmental factors. Multiple sensors are integrated with an Arduino Mega 2560 to detect important environmental factors such as sunlight intensity, temperature, humidity, wind speed, rainfall, atmospheric pressure, voltage and current. These data were logged to an SD card every 5 minutes over two months period and pre-processed for the training. The set of data is then imported to MATLAB to train NARX model using the Levenberg-Marquardt algorithm. This model achieved a best validation performance of 0.00035075 MSE, showing a high accuracy in analysing the non-linear relationships between environmental conditions and solar power output. The time series graph also shows that the predicted values are closer to the actual values during training, validation and testing stages. This forecasting method highlights the importance of using the NARX model in improving the solar energy under varying weather conditions.