In commercial microgrids, for effective energy management and reliable decision-making, it is imperative to include the uncertainties in load demands as well as in the renewable energy generation. To refine the research focus, the proposed work focusses on solar PV-load forecasting in the smart grid environments. The rise in demand fluctuations necessitates an improved 10% accuracy in the forecasting models. But whenever the forecasting horizon length crosses time steps of 12, the existing models becomes unstable in predicting, with a degradation in accuracy and scalability by 8-15%. Limited research works are available to analyze the effect of rolling-horizon in managing the uncertainties. In this scenario, the proposed work employing transformer-based PV-load forecasting framework, can achieve an improved probabilistic forecasting accuracy of greater than 12%. Horizon-aware learning mechanisms are incorporated into the proposed model to accurately estimate the uncertainty. Model rolling-horizon based experiments using MATLAB environment simulation is performed on the proposed model for validation purposes. All forecasts and probability analyses illustrated in the present document are produced through simulation results by the authors using consistent simulation conditions. Three different operational conditions are considered for the evaluation with the performance metrics being the continuous ranked probability score (CRPS) and pinball loss. Improved quality of probabilistic forecasting is visible with a reduction in CRPS value by 12.6% and effective prediction interval capture is seen with internal coverage of 9.4%. The computational cost and requirements have been lowered by 18%. The architecture can scale up to about 14 different forecasting horizons, with consistent stability under different PV and load conditions. The numerical results confirm consistent improvements in the performance gains of the proposed forecasting model. Further this transformer-model based approach outperforms both gates recurrent unit baselines and long short-term memory models. There is a 12% gain improvement in average case scenarios and about 6% gain improvement in worst case scenarios, making the model suitable for applications that demand latency time of less than 1.5 s. Thus, the detailed analysis demonstrates performance gains in terms of different evaluation metrics. Thus, the overall results exhibit superior performance as against other existing modelling techniques. When scaling up the transformer-based modelling concept beyond horizon steps of 14, there is a degradation in the forecasting performance, which can be analyzed in future scope.
Deepa et al. (Tue,) studied this question.