Search

This paper aims to provide a thorough guide on how to analyze complex energy demand forecasting models with Shapley Additive exPlanations (SHAP) in order to build trust in their predictions and understand the model and SHAP limitations based on selected real-world use case. There are only few examples described in the literature in energy industry and they present very basic usage. This study fills the gap for the class of energy (heat, electric, gas) demand predicting models by showing step by step, top-down analysis of state-of-the-art, deep neural network model predicting total heat demand (hot water and room heating) in Warsaw District Heating Network – the largest district heating network in EU. The paper shows how SHAP can be successfully used on demand forecasting models to provide practical, easily interpretable insights on inner workings of these models which can be used to assess their reliability and plan further development. • Predictions of deep neural network models forecasting energy demand can be explained. • SHAP can be successfully used to analyze deep ANN with multiple inputs (>100). • Step by step guide on how to analyze energy demand model with SHAP is provided. • Relationships identified by ANN forecasting heat demand agree with expert knowledge. • Trust in predictions of black-box models can be build by analyzing with SHAP. [ABSTRACT FROM AUTHOR]

This paper presents a complex analysis of heat demand forecasting methods for the Warsaw District Heating Network, which is owned by Veolia Energia Warszawa, the largest district heating network (DHN) in the European Union. The analyzed network supplies heat for both domestic and heating purposes. Therefore, summer, intermediate, and winter seasons were delineated and separately evaluated. Numerous models were utilized including models broadly recognized and used (ridge regression, autoregression with exogenous input, deep artificial neural networks), as well as previously unexplored models (combination of summer and winter linear models with the utilization of fuzzy logic). A 72 h forecast horizon is evaluated for total heat demand (the sum of all substations), as well as for groups of buildings (local models for specific city areas), and individually for the majority of substations. Models of areas use an additional input variable, namely, the results of the total heat demand forecast, and are proposed to be developed as an auxiliary information variable offered to grid operators. An artificial neural network based model achieves the best accuracy for all analyzed seasons. The intermediate seasons prove to be the most difficult to accurately forecast for and only the combination of summer and winter linear autoregresive models with utilization of a fuzzy logic reached comparable accuracy. • Heat demand for district heating network is calculated. • A 24 h horizon heat demand forecasts are proposed and evaluated. • Artificial neural network for total heat demand presents the best accuracy. • Ridge regression and fuzzy logic for network areas presents the best accuracy. • Implementation in on-line system is presented. [ABSTRACT FROM AUTHOR]