Browsing by Author "Mbuya, Benson"

Now showing 1 - 2 of 2
  • Thumbnail Image
    Item
    Short-term forecast techniques for energy management systems in microgrid applications
    (NM-AIST, 2023-07) Mbuya, Benson
    In the 2015 Paris Agreement, 195 countries adopted a global climate agreement to limit the global average temperature rise to less than 2°C. Achieving the set targets involves increasing energy efficiency and embracing cleaner energy solutions. Although advances in computing and Internet of Things (IoT) technologies have been made, there is limited scientific research work in this arena that tackles the challenges of implementing low-cost IoT-based Energy Management System (EMS) with energy forecast and user engagement for adoption by a layman both in off-grid or microgrid tied to a weak grid. This study proposes an EMS approach for short-term forecast and monitoring for hybrid microgrids in emerging countries. This is done by addressing typical submodules of EMS namely: load forecast, blackout forecast, and energy monitoring module. A short-term load forecast model framework consisting of a hybrid feature selection and prediction model was developed. Prediction error performance evaluation of the developed model was done by varying input predictors and using the principal subset features to perform supervised training of 20 different conventional prediction models and their hybrid variants. The proposed principal k-features subset union approach registered low error performance values than standard feature selection methods when it was used with the ‘linear Support Vector Machine (SVM)’ prediction model for load forecast. The hybrid regression model formed from a fusion of the best 2 models (‘linearSVM’ and ‘cubicSVM’) showed improved prediction performance than the individual regression models with a reduction in Mean Absolute Error (MAE) by 5.4%. In the case of the EMS blackout prediction aspect, a hybrid Adaptive Similar Day (ASD) and Random Forest (RF) model for short-term power outage prediction was proposed that predicted accurately almost half of the blackouts (49.16%), thereby performing slightly better than the stand-alone RF (32.23%), and ASD (46.57%) models. Additionally, a low-cost EMS smart meter was developed to realize the implemented energy forecast and offer user engagement through monitoring and control of the microgrid towards the goal of increasing energy efficiency.
  • Thumbnail Image
    Item
    Short-term load forecasting in a hybrid microgrid: a case study in Tanzania
    (Journal of Electrical Systems, 2019) Mbuya, Benson; Moncecchi, Matteo; Merlo, Marco; Kivevele, Thomas
    Most emerging countries such as Tanzania are promoting rural electrification through installation of microgrids. This paper proposes an approach for short-term day-ahead load forecast in rural hybrid microgrids in emerging countries. Energy4Growing research project by Politecnico di Milano department of energy in collaboration with EKOENERGY (www.ekoenergy.org) implemented in Ngarenanyuki Secondary School (Arusha, Tanzania) innovative control switchboards to form an energy smart-hub. The smart-hub was designed to manage the school’s 10kW hybrid micro-grid comprising: PV-inverter, battery storage, microhydro system, and genset. Ngarenanyuki school microgrid’s data was used for the experimental short-term load forecast in this case study. A short-term load forecast model framework consisting of hybrid feature selection and prediction model was developed using MATLAB© environment. Prediction error performance evaluation of the developed model was done by varying input predictors and using the principal subset features to perform supervised training of 20 different conventional prediction models and their hybrid variants. The objective function was feature minimization and error performance optimization. The experimental and comparative day-ahead load forecast analysis performed showed the importance of using different feature selection algorithms and formation of hybrid prediction models approach to optimize overall prediction error performance. The proposed principal k-features subset union approach registered low error performance values than standard feature selection methods when it was used with ‘linearSVM’ prediction model. Furthermore, a hybrid prediction model formed from the elementwise maximum forecast instances of two regression models (‘linearSVM’ and ‘cubicSVM’) yielded better MAE prediction error than the individual regression models fused to form the hybrid.