Intelligent energy management system for optimum reserve allocation and multiobjective demand response using multi-threaded SVR-LSTM model

Abstract

Demand response is a critical standard for any power system. Prospective demand forecasting enables a smooth demand response with better system stability and optimum reserve allocation. This project has developed an expert energy management system that works by monitoring, analyzing, and controlling elements of the power system for quick and optimum demand response. The hardware system monitors each component of the power system and this information is sent to a central server using a low-power wide-area network (LPWAN). With this data, successful forecasting of the demand is achieved with the help of a Long Short-Term Memory neural network model (LSTM). LSTM has greater accuracy, but comparatively, high computing time compared to other time series methods such as ARIMA, SARIMA, RNN, etc. The accuracy of LSTM is compromised when issues like exploding gradient, vanishing gradient, uniform gradient, etc. are developed. These issues are addressed in this paper and a new model is developed based on a Support vector machine-based multi-threaded multi-layer LSTM. Extensive metadata analysis is conducted for accurate sampling and filtering. Support machine vector regression algorithms are integrated through multiple parallel threads to eliminate gradient errors, thus increasing precision. In the event of any gradient errors, SVR can reallocate the weight distribution in an adaptive approach based on historical data. The SVR also helps to reduce computation time and error by adjusting the epoch level in sub-threads based on the time-series variance supported by the metadata. A fully polynomial multi-purpose time approximation method is used for the allocation of reserves, primarily according to the forecasted values and the constraints of the power system such as diversity factor, load factor, etc. The functionality of the system is improved by the adaptive redesign of the model with the help of multi-threading and variable forget gate. The testing and real-time implementation of the model with the Hardware in a Loop (HIL) system is successfully carried out. The accuracy of the system is ensured with standard RMSE analysis. Based on the error analysis, it is clear that the error up to 200 % reported in the conventional method was reduced to 0.01 % using the adaptive SVR-LSTM method. The computing time is also reduced by half or less than the conventional system. The distributed control method is used for the individual monitoring and control.