Forecasting Energy Demand For Microgrids Over Multiple Horizons Fred Otieno 1 , Nathan Williams 1 , Patrick McSharry 123 1 Carnegie Mellon University Africa, Kigali, Rwanda 2 African Centre of Excellence in Data Science, University of Rwanda, Kigali, Rwanda 3 Oxford Man Institute of Quantitative Finance, University of Oxford, Oxford, UK Abstract—Access to electricity is one of the key enablers of socio- economic development in Sub-Saharan Africa. Microgrid solutions are currently playing an increasing role in providing access to electricity, especially to rural populations whose electricity is not supplied by the national grid. Microgrid developers need to manage their existing sites and expand to new regions. In order for them to manage this expansion effectively and sustainably, they need to make data-driven decisions. Having access to accurate forecasts of electricity demand at the site level is a key input in designing, managing and up-scaling microgrid solutions. Several forecasting mechanisms are proposed for such microgrid developers. Using daily energy consumption data from seven sites operating in Kenya during 2014-2017, it was established that exponential smoothing offers the best out-of- sample forecasting performance with forecast skill exhibited for horizons up to four months ahead. Index Terms--Demand forecasting, Microgrids, Power demand. I. INTRODUCTION In Sub-Saharan Africa, nearly 600 million people - about 70% of the population - live without electricity [1]. The International Energy Agency foresees that microgrids and other off-grid solutions will have a huge role in providing energy for 70% of all rural populations in developing countries [2]. This has led to the embracing of off-grid and distributed energy approaches. In Kenya, the electrification rate was 36% in 2014 [3]. Access to electricity is crucial for driving economic development in Africa [4]. Therefore, there is a pressing need to increase the number of people that have access to electricity. Privately owned microgrids are one of many solutions to increasing electricity access in Kenya. Grid extension, public and community microgrids, and stand-alone systems all have a role to play. Private microgrid developers are likely to first target sites and customers viewed as having the highest potential for electricity consumption in order to achieve a commercial return on investment. Without subsidies and other support mechanisms, microgrid rollout and market penetration will be limited to these areas. Nevertheless, overall, private microgrids can provide an important contribution to off-grid electrification efforts for the rural market in Kenya. The challenge for these offgrid providers is the ability to design systems that will be able to meet the demand of its customers over time. One such offgrid company is PowerGen Renewable Energy. PowerGen supplies energy solutions to community, home, business and light industrial clients in East Africa. Founded in 2011, PowerGen has set up over 40 microgrids and currently serves thousands of customers across seven countries with clean, renewable energy. It offers system design and engineering, device and technology procuring, implementation and integration, and operations and maintenance services. It is redefining rural electrification in Kenya using these microgrid solutions. As PowerGen seeks to expand its operations, a challenge they face is customer acquisition and demand assessment. Demand assessment can be conducted through forecasting. This will enable the company to venture into new sites once economic viability has been established. Apart from setting up new sites, this research aims to provide an automated forecasting system for PowerGen. The focus in primarily to facilitate performance monitoring of a site that has a microgrid deployed. This will enable PowerGen to preempt future peaks and troughs in demand and enable the system to be modified accordingly. An efficient load forecasting model will serve PowerGen with reliability in terms of scheduling, planning and managing their microgrid. According to McSharry et al. there are various challenges associated with different load forecasting horizons [5]: • Very short-term load forecasting. This ranges from seconds and minutes to several hours. This is required for controlling the flow. • Short-term load forecasting. Ranges from hours to weeks. These forecasts are useful in adjusting generation and demand and informing launch of offers to the electricity supply market. • Medium-term and long-term load forecasting. Ranges from months to years. These forecasts are normally used to plan power generation asset utilities. The crucial forecasting horizons are weekly, daily and hourly. These horizons have a direct impact on the day to day operation of the power generation company. Being able to foresee an upcoming demand spike can enable the power generation company to adjust accordingly in order to meet the needs of their customers and maintain the reliability of service. Based on the different forecasting horizons, various forecasting methods have been recommended by Hernandez et al. [6]. Due to the stochastic nature of electricity demand as a function of time the seasonal ARIMA and state space models 2018 IEEE PES/IAS PowerAfrica 457 978-1-5386-4163-7/18/$31.00 ©2018 IEEE