An Efficient SPV Power Forecasting using Hybrid Wavelet and Genetic Algorithm based LSTM Deep Learning Model

Solar photovoltaics (SPV) are widely favoured energy generation system that has seen its rapidly growing installed capacity in the power system structure from past few decades. This grows concern of operation and control due to its high stochastic nature and dependence on weather variables such as temperature, irradiance, humidity etc,. This makes SPV power forecasting necessary in order to manage and plan to get useful insights. This article proposes a hybrid wavelet and genetic algorithm (GA)-based long short term memory (LSTM) deep neural network model to forecast SPV output power of a 58 MW utility scale SPV plant installed in the Florida state. The data are obtained from publicly available NREL database with 5-min resolution. Temperature and relative humidity along with historical SPV output power has been used as input features to the neural network model. Discrete wavelet transform is applied in order to denoise the data and due to its inconstancy, which increases the data dimension and helps in improving forecasting accuracy. GA has been combined with LSTM to find the optimized window size and LSTM units. The proposed method is then compared with different benchmark methods such as persistent/naive, state vector regression (SVR) and long short term memory-deep neural network (LSTM-DNN) model structure. The results shows an improvement of accuracy in terms of performance metrics most commonly used in machine learning such as mean squared error, root mean squared error, mean absolute error and r-squared values.