978-1-6654-4389-0/21/$31.00 ©2021 IEEE Micro-Genetic Algorithm Embedded Multi- Population Differential Evolution based Neural Network for Short-Term Load Forecasting Colin Paul Joy School of computing, Engineering and Digital technologies Teesside University Middlesbrough, UK c.joy@tees.ac.uk Dr Gobind Pillai School of computing, Engineering and Digital technologies Teesside University Middlesbrough, UK g.g.pillai@tees.ac.uk Dr Kamlesh Mistry Dept. of computer and information sciences Northumbria University Newcastle Upon Tyne, UK k.mistry@northumbria.ac.uk Dr Yingke chen School of computing, Engineering and Digital technologies Teesside University Middlesbrough, UK y.chen@tees.ac.uk Abstract—The load of a power system usually presents a certain range of nonlinear fluctuation with time. Even then, the load characteristics still follow certain rules which can be exploited to optimise and improve the accuracy of computer- based Short-Term Load Forecasting (STLF) models. Therefore, this paper presents a mGA (micro–Genetic Algorithm) embedded multi-population DE (Differential Evolution) to optimise an Artificial Neural Network (ANN) STLF model. Firstly, the mGA embedded multi-population DE is proposed, to improve and balance the global and local search. Then the proposed DE is applied to optimise the weights during the training of the ANN. The overall model’s performance is evaluated using publicly available Panama electricity load dataset against four state-of-the-art machine learning algorithms. The evaluation results show that the proposed DE based NN STLF model has higher prediction accuracy compared to the other selected machine learning algorithms. Keywords—Machine Learning, Differential Evolution, Neural Networks, Electricity Load Forecasting I. INTRODUCTION The process of predicting the demand for electricity consumption during a specific time frame in a specific area is referred to as Load forecasting. Load forecasting is an effective and important technique that assists in the management and operation of power systems which can contribute to significant reductions in costs when conducted precisely [1]. Currently, electric storage such as batteries and pumped hydro storages have limited capacities. Because of the lack of mechanisms to store electricity in large quantities, at the national scale, the amount of electricity that is produced at a given time has to cover the demand and compensate for the losses. Hence, there is always a requirement to track the load by system generation for effective operation of a power system. The forecasting carried out for a single day to several weeks ahead is usually referred to as short-term load forecasting (STLF). This process supervises the production of electricity in power stations [2] by tackling electricity wastage, manpower wastage and assist in mapping power infrastructure. Faultless and precise load prediction has forever been a continuous challenge in research since if the predicted load is lesser than required load there will arise a shortage of electricity. Meanwhile, if the predicted load is surplus to the required load this will lead to wastage of energy, machine resources and drive up the costs. Load forecasting helps to establish balance in regional power supply and demand while simultaneously making sure the whole system is run safely. In order to build the load prediction models in this paper, the city of Panama is chosen as it shares its electricity load data to the public providing us with much needed research data considering most collective and multi-site electricity demand statistics are difficult to obtain. Based on methodologies proposed by the National Dispatch Centre (CND) [3], the organization that is in charge of power system planning and operations in Panama, the final aim of forecasting with minimum deviation is to predict and supply the demand while incurring minimum losses. The CND performs the forecast planning weekly and ensures short term forecasting for the following week is efficient enough to tackle security issues in the electrical system. To conduct short term scheduling, the CND utilizes an optimization software that handles input hourly [5] by scrutinizing data related to load forecast, power plants and the power grid. The Nostradamus Artificial Neural Network (ANN) by Hitachi ABB [6] is employed by the National Dispatch Centre for optimization helping in hourly and weekly predictions [7]. Fig. 1. System Architecture diagram