Machine Learning-Based Power Output and Load Prediction for a Solar Photo-voltaic Mini-grid

Abstract

The power output of solar photovoltaic (PV) mini-grids is greatly affected by the dynamically changing weather conditions and load demand. This affects power production planning during the operation of these systems. In order to improve power production planning, there is need for approaches that accurately predict the power output and load demand for PV mini-grids.

In this project, novel machine learning approaches are proposed for next hour PV power output and load demand prediction. The proposed methods take as input time series data, consisting of weather and schedule variables, and utilize long short-term neural networks (LSTMs) that are capable of modelling temporal dependencies between the data and the PV power output.

The proposed methods achieve state-of-the-art performance on open datasets with the power output prediction model achieving a normalized root mean square error (RMSE) of 0.092 and the load prediction model achieving a normalized RMSE of 0.0855. The power output prediction model was also evaluated on a dataset generated by simulating the Utility 2.0 solar PV mini-grid in Kiwumu, Mukono Uganda. The model achieves similar results as those obtained on the open dataset.