• A novel forecasting method utilizing AI was developed to predict power demand for both the research laboratory building and the entire Higashi-Hiroshima campus of Hiroshima University.
• The method incorporates Bayesian inference, which probabilistically estimates demand by integrating prior knowledge with observed data.
• Validation using actual demand data demonstrated that the proposed method achieves prediction accuracy comparable to or better than conventional models. Additionally, it supports explainability, quantification of uncertainty, and is suitable for real-time processing.

In recent years, regional microgrids have attracted attention as a future vision for electricity and energy supply systems that enhance resilience against natural disasters. At the core of these systems lies the Energy Management System (EMS), which has been actively researched by various institutions.

Within an EMS, accurately capturing and forecasting electricity demand is essential for properly managing the balance between energy supply and consumption. The accuracy of these forecasts significantly affects the efficiency and stability of the system.

In this study, we developed a demand forecasting method by setting the following three requirements necessary for practical EMS operation, in addition to forecast accuracy:
　・The rationale behind the forecast must be explainable.
　・The uncertainty associated with the forecast must be quantifiable.
　・The computation must be completed within the EMS update cycle.

To meet these requirements, we aimed to mimic the sequential and intuitive forecasting thought process of humans, which combines heuristics and situational judgment. For this purpose, we adopted a linear regression model based on Bayesian inference, which allows flexible integration of past experiences and new information.

Although Bayesian linear regression models are generally considered unsuitable for electricity demand forecasting due to inherent nonlinearity, we improved forecast accuracy through innovations in data collection and a unique weighting scheme based on temporal proximity. Validation using real data from the research buildings and the entire campus of Hiroshima University demonstrated forecast accuracy equal to or better than existing models such as LightGBM and LSTM, which are known for their strength in handling nonlinearity.

Furthermore, the proposed method satisfies the three practical requirements and can flexibly adapt to various scales of demand, making it a promising technological foundation for advancing EMS in regional microgrids.