Abstract:
The uncertain nature of photovoltaic generation is a challenge in energy
management and may occur a mismatch between load and power generation notably
in the grids with high levels of solar energy penetration, so the development of PV
generation forecasting methods has received attention recently. In this research, short term PV power generation has been forecasted using MATLAB software according to
the past generation trend by the three hybrid models. Multilayer perceptron (MLP)
neural network (NN) was implemented as the predictive network and trained by
optimization algorithms for achieving the model’s optimal performance by
determining the hyperparameters of the MLP NN which are the input and hidden layer
weights matrixes and biases vector. For the purpose of minimizing the error between
actual and forecast values, RMSE was considered as the fitness function. In the first
model, particle swarm optimization (PSO) is employed as an inner block and it tunes
the hyperparameters of the network. Imperialist competitive algorithm (ICA) and
genetic algorithm (GA) are the network optimizers in the second and third models
respectively. PV generation data in July and August 2022 (15-minute intervals) in
Belgium is considered as the data set. The assessment metrics are six different errors
includes Mean Square Error (MSE), Root Mean Square Error (RMSE), Mean Absolute
Error (MAE), Mean Bias Error (MBE), Coefficient of determination (R2
), and
normalized RMSE (NRMSE).
The results express that the combination of the MLP NN and PSO has better
performance. Fewer errors in this hybrid model and the higher value of R2
express that
this hybrid model has more accurate output, and acceptable accuracy in time series
forecasting.
