Solar Irradiance Forecasting using Hybrid Long-Short-Term-Memory based Recurrent Ensemble Deep Random Vector Functional Link Network

Accurate and reliable forecasting of solar irradiance is necessary for an efficient grid performance with large scale penetration of photovoltaic (PV) generation. Thus, with an aim to improve solar irradiance forecasting accuracy, a new decomposition based hybrid model known as Stacked Long-Short-Term-Memory (LSTM) recurrent neural network is proposed in this paper. Further the dense layer of the stacked LSTM architecture is replaced by a novel Recurrent Ensemble Deep Random Vector Functional Link Network (REDRVFLN) to improve generalisation, speed up computation, and prediction accuracy. The raw irradiance data is pre-processed using Isolation Forest (IF) algorithm to remove the presence of outliers from the data and the Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) algorithm decomposes the preprocessed data into Intrinsic Mode Functions (IMFs) with zero reconstruction error and better separation of spectral components. The recurrent stacked LSTM neural network effectively captures the temporal features and long term dependencies of decomposed solar irradiance time series data. On the other hand REDRVFLN model comprising several stacked layers of locally recurrent neurons with fixed random weights and biases effectively handles processed temporal features from the LSTM module with optimal generalisation and improved stability. Further the ensemble of the outputs from each layer produces the final forecast with better accuracy in comparison to many widely used deep neural network and other benchmark models. The performance of the proposed stacked LSTM integrated REDRVFLN model has been validated using solar irradiance data samples both hourly and with seasonal variations producing superior accuracy.