Design Support Using a Neural Network Algorithm

General Introduction

The summary of the 1st section is as follows. A microgrid with the capacity for sustainable energy is expected to be a distributed energy system that exhibits quite a small environmental impact. In an independent microgrid, “green energy,” which is typically thought of as unstable, can be utilized effectively by introducing a battery. In the past study, the production-of-electricity prediction algorithm (PAS) of the solar cell was developed. In PAS, a layered neural network is made to learn based on past weather data and the operation plan of the compound system of a solar cell and other energy systems was examined using this prediction algorithm. In this study, a dynamic operational scheduling algorithm is developed using a neural network (PAS) and a genetic algorithm (GA) to provide predictions for solar cell power output. We also do a case study analysis in which we use this algorithm to plan the operation of a system that connects nine houses in Sapporo to a microgrid composed of power equipment and a polycrystalline silicon solar cell. In this work, the relationship between the accuracy of output prediction of the solar cell and the operation plan of the microgrid was clarified. Moreover, we found that operating the microgrid according to the plan derived with PAS was far superior, in terms of equipment hours of operation, to that using past average weather data.

The summary of the 2nd section is as follows. The bioethanol reforming system (FBSR) using sunlight as a heat source is a fuel production system for fuel cells with little environmental impact. However, because solar radiation and outside air temperature are unstable, it is difficult to predict operation of the system with accuracy. Therefore, an operation prediction program of the FBSR using a layered neural network (NN) with the error-correction learning method has been developed. We developed a method of analyzing the operation of a natural energy system with sufficient accuracy. The weather pattern (the amount of global solar radiation and the outside air temperature) and energy-demand pattern for the past one year are inputted into the NN. Moreover, training signals are calculated by a genetic algorithm (GA). The training signals are given to the NN, and the operation pattern of the FBSR is made to learn. Operation of the FBSR on arbitrary days can be predicted by inputting the weather pattern and the energy-demand pattern into this learning NN. In this study, the operation prediction program of the FBSR is developed, and details of the analytic accuracy are clarified. As a result of analyzing using the developed algorithm, when ±20% or less of power load fluctuation occurred, the operation plan was analyzable in 14% or less of error span. On the other hand, in operation prediction when ±50% or less of fluctuation is added to the outside temperature and global solar radiation, there was 16% or less analysis error.