End-to-End Forecasting Towards Economic Operation of Microgrid Using Derivative-Free Learning

The economic operation of microgrid (MG) relies on day-ahead forecasts to manage uncertainties from renewable energy sources (RES) and load demands. However, precise fore-casts do not necessarily guarantee reduced operational costs. This paper introduces an integrated end-to-end forecasting (E2EF) approach that utilizes deep neural networks (DNNs) to merge RES and load forecasting tasks with MG dispatch challenges. The E2EF's effectiveness is evaluated based on it's impact on MG dispatching costs. To optimize the forecast model within the end-to-end framework, two critical techniques are introduced: a genetic algorithm (GA) employed as a derivative-free technique for end-to-end optimization, and a pre-trained accuracy-focused model for warm-start. The first technique averts the negative impacts of steep gradients that commonly impede optimization within E2EF, while the second ensures GA improves upon the benchmark set by the well-trained accuracy-focused model, minimizing the inefficiencies of heuristic approaches. Case studies using real-world datasets reveal that the proposed E2EF methodology markedly reduces the operational costs of MG compared to conventional accuracy-focused forecast model, while also outperforming existing end-to-end approaches.