PHOTOVOLTAIC POWER RAMP EVENT FREQUENCY FORECASTING USING A HYBRID DEEP LEARNING NEURAL NETWORK

Abstract

One major concern in the solar power industry is the controlling of abrupt changes in power output in order to optimise the process of electricity generation from solar panels. In this dissertation, an existing Convolutional Neural Network - Long Short-Term Memory (CNN-LSTM) deep learning model is studied and experimented on the performance in photovoltaic power ramp event frequency prediction with the assistance of the Swinging Door Algorithm (SDA) algorithm. A quantitative formulation of the requirements that a ramp event satisfies is also provided. In addition, 2 other models are also assessed along with the CNN-LSTM model to provide comparison for more insights and a better picture of the capability of the CNN-LSTM model. The data used for training and testing are collected from the power consumption data of 5 houses in the UK, with several steps of pre-processing to fit the purpose of the experiment. The testing results are evaluated using 3 evaluation metrics named MAE, RMSE and R-Squared Score. Experiment data shows that the hybrid CNN-LSTM model performs more accurately than the other 2 models in terms of predicting the number of possible ramps for each set of 4-hour power output data provided. A self-implemented SDA with some modifications is also presented in this dissertation. Moreover, this dissertation hopes to encourage the study of applying deep learning techniques in wind power into photovoltaic power.