Comparison of Data Partitioning Method using LSTM for Power Demand at a Clinic

In Japan, securing power to maintain medical services at hospitals and clinics is an important issue in the event of a power outage during a disaster such as a typhoon or earthquake. Systems that combine photovoltaic power generation, storage batteries, and diesel generators (DGs) are considered as power systems for hospitals and clinics. Operating power systems with optimal fuel storage capacity for DGs and charging capacity of storage batteries necessitate accuracy of power demand forecasting. Previous studies on power demand forecasting methods have considered the structure of predictors and neural networks used for forecasting models. However, the data period of the predictors, data partitioning methods, and the hidden layer have seldom been considered. Therefore, in this paper, the data period, data partitioning methods, and number of nodes in the hidden layer are considered as parameters of forecasting models using long short-term memory. Results show that learning results can be improved by increasing the number of nodes in the case of block partitioning. Furthermore, the data period, which has lengths that allows learning the periodicity of data indicates lower mean absolute error (MAE) and root mean square error (RMSE) values than random partitioning. The model with a data period of 2 years, 200 nodes and block partitioning indicate a 12% improvement in MAE and a 9.1% improvement in RMSE. In the case of random partitioning, the influence of the number of nodes and data period can be reduced, and if there is not enough data to learn the periodicity of data, the learning results can be improved by using random partitioning. The model with data period of 1 year, the number of nodes is 100, and random partitioning denote a 29.7% improvement in MAE and a 30.3% improvement in RMSE.