This paper proposes a novel forecasting model designed to accurately forecast the PV power output for both large‐scale and small‐scale PV systems. The proposed model uses available temperature data, approximate and detailed coefficients obtained from the decomposed PV power time series using the stationary wavelet transform (SWT), and statistical features extracted from the historical PV data. The model is comprised of four long–short–term memory (LSTM) recurrent neural networks (RNN) designed to perform multi‐step forecasting on the individual approximate and detailed coefficients decomposed by the SWT and a final deep neural network (DNN) designed to perform the next time step PV power forecast. The DNN makes use of the reconstructed values estimated by the four LSTM networks together with temperature data and statistical features to predict the final forecasted value of the next time step PV power. 30‐min resolution data from a 12.6 MW PV system located in the state of Florida are used for testing and evaluating the proposed method against several models found in the literature. The results obtained suggest that the proposed model improved the forecasting accuracy significantly in the metrics used to compare with other models while reducing the number of features needed to perform the forecasting operation.