Effect of Short-term and High-resolution Load Forecasting Errors on Microgrid Operation Costs

Abstract

The aim of this paper is to evaluate the effect of the load forecasting errors to the operation costs of a gridconnected microgrid. To this end, a microgrid energy scheduling optimization model was tested with deterministic and stochastic formulations under two solution approaches i.e., day-ahead and rolling horizon optimization. In total, twelve simulation test cases were designed receiving as input the forecasts provided by one of the three implemented machine learning models: linear regression, artificial neural network with backpropagation, and long short-term memory. Simulation results of the weekly operation of a real residential building (HSB Living Lab) showed no significant differences among the costs of the test cases for a daily mean absolute percentage forecast error of about 12%. These results suggest that operators of similar microgrid systems could use simplifying approaches, such as day-ahead deterministic optimization, and forecasts of similar, non-negligible accuracy without substantially affecting the microgrid's total cost as compared to the ideal case of perfect forecast. Improving the accuracy would mainly reduce the microgrid's peak power cost as shown by its 20.2% increase in comparison to the ideal case.

Water Wave Optimization Algorithm-Based Dynamic Optimal Dispatch Considering a Day-Ahead Load Forecasting in a Microgrid

“…Several techniques are commonly used to forecast dayahead load in the microgrid such as time series analysis (TSA) [27]; decision trees and random forests (DTRF) [28]; gradient boosting model (GBM) [29]; neural network (NN) with training algorithms of Levenberg-Marquardt, Bayesian Regularization, and Scaled Conjugate Gradient [30]- [31]; support vector regression (SVR) [32]; convolutional neural network (CNN) [33]; long short-term memory (LSTM) [34]; Kalman filtering (KF) [35]; and linear regression (LR) [36].…”

“…The comparison of these techniques used to forecast dayahead load in the microgrid is shown in Table II through various specific features such as properties of input data, management of non-linearity, management of missing data, complexity in model training, management of multicollinearity, management of outliers, adaptability to real-time, and adaptability to large datasets [27]- [36]. Amongst the above techniques, the CNN and LSTM techniques are more prominent than others.…”