This work is licensed under a Creative Commons Attribution 3.0 License. For more information, see http://creativecommons.org/licenses/by/3.0/. This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2019.2923796, IEEE Access VOLUME XX, 2017 1 Date of publication xxxx 00, 0000, date of current version xxxx 00, 0000. Digital Object Identifier 10.1109/ACCESS.2017.Doi Number Improving Load Forecasting Process for a Power Distribution Network Using Hybrid AI and Deep Learning Algorithms Sibonelo Motepe 1 , Ali N. Hasan 1 , Member, IEEE, and Riaan Stopforth 2 , Senior Member, IEEE 1 Faculty of Engineering and the Built Environment, University of Johannesburg, Johannesburg 2092, South Africa 2 Stopforth Mechatronics Robotics Research Lab, School of Engineering, University of Kwa-Zulu Natal, Durban 4041, South Africa Corresponding author: Sibonelo Motepe (e-mail: djscvii@gmail.com). The authors would like to thank the South African Weather Services for providing them with weather data. The authors also acknowledge the National Research Foundation, the Eskom TESP programme, and the DST ROSSA programme for partially funding this research. ABSTRACT Load forecasting is useful for various applications including maintenance planning. The study of load forecasting using recent state-of-the-art hybrid artificial intelligence (AI) and deep learning (DL) techniques is limited in South Africa (SA) and South African power distribution networks. This paper proposes a novel hybrid AI and DL South African distribution network load forecasting system. The system comprises of modules that handle the collection of the loading data from the field, analysis of data integrity using fuzzy logic, data preprocessing, consolidation of the loading and the temperature data, and load forecasting. The load forecasting results are then used to inform maintenance planning. The load forecasting is conducted using a hybrid AI/DL load forecasting module. A novel comparative study of recent state of the art AI techniques is also presented to determine the best technique to deploy in this module when forecasting South African power redistributing customers’ loads. The impact of the inclusion of weather parameters and loading data clean up on the load forecasting performance of a hybrid AI technique, optimally pruned extreme learning machines (OP-ELM), and a deep learning technique, long short-term memory (LSTM), is also investigated. These techniques are compared with each other and also with a commonly used powerful hybrid AI technique, adaptive neuro-fuzzy inference system (ANFIS). LSTM was found to achieve higher load forecasting accuracies than ANFIS and OP-ELM in forecasting the two distribution customers’ loads in this study. Only LSTM models’ performance improved with the inclusion of temperature in their development. INDEX TERMS Adaptive Neuro-Fuzzy Inference Systems, Artificial Intelligence, Deep Learning, Distribution Networks, Extreme Learning Machines, Load Forecasting, Recurrent Neural Networks, Long Short-Term Memory I. INTRODUCTION Electricity has been regarded as South Africa’s gross domestic product’s (GDP) main driver [1], [2]. Developing countries still experience a lack of electricity access [3]-[6]. These countries, including South Africa, have electrification programs that are driving the connection of its citizens to the power grid. South Africa (S.A.) obtained its democracy in 1994, and has since then electrified more than 5.2 million homes and over 12 000 schools [7]. The South African government plans to achieve universal supply by 2025/2026 [8]. In order to achieve this goal, while ensuring continuity of supply, utilities need planning at different levels of the power system. Load forecasting whose importance was established in different studies including [9] and [10], becomes important in order to achieve a sustainable power supply. Load forecasting has different windows which it can be classified into. These windows are short term, medium term and long term, which respectively cover hours to weekly forecasts, monthly to quarterly forecasts and then yearly forecasts [11]. With the movement towards the smart grid in developed countries, recent load forecasting studies have moved past the customer supply point [12]-[14]. Appliance power consumption data have been incorporated to forecast