Thanks to their modularity and their capacity to adapt to different contexts, hybrid microgrids are a promising solution to decrease greenhouse gas emissions worldwide. To properly assess their impact in different settings at country or cross-country level, microgrids must be designed for each particular situation, which leads to computationally intractable problems. To tackle this issue, a methodology is proposed to create surrogate models using machine learning techniques and a database of microgrids. The selected regression model is based on Gaussian Processes and allows to drastically decrease the computation time relative to the optimal deployment of the technology. The results indicate that the proposed methodology can accurately predict key optimization variables for the design of the microgrid system. The regression models are especially well suited to estimate the net present cost and the levelized cost of electricity (R2 = 0.99 and 0.98). Their accuracy is lower when predicting internal system variables such as installed capacities of PV and batteries (R2 = 0.92 and 0.86). A least-cost path towards 100% electrification coverage for the Bolivian lowlands mid-size communities is finally computed, demonstrating the usability and computational efficiency of the proposed framework.@en

With pressing priorities in the development agenda, policy makers in developing countries are in the difficult situation of prioritizing policy actions. Limited government and utility budgets need cost effective solutions to bring the desired development benefits of electrification, health, education and food security among others. Energy access is a prerequisite for economic activity and for human development as interacts in synergy with other development needs. As rural electrification models usually focus on the supply of electricity solely, thermal energy needs, such as cooking and water heating remain unattended and satisfied by non-renewable energy fuels. To this aim, we explore optimal electrification solutions addressing two types of energy demands, electricity and thermal energy demands for cooking. Our model builds on a 3-step electrification methodology proposed by Peña et al. including electricity as a modern source of clean energy for cooking in rural communities. The total investments needed to build and operate the microgrids, including distribution costs, is 332 million USD. This is equivalent to 1129 USD/per inhabitant. This amount does not account however the health and environment benefits that e-cooking can bring to inhabitants in Bolivian low-lands.@en

Countries around the world are preparing to give the last leap to accomplish a 100 % of rural energy access. Nonetheless, country-wide electrification planning requires the analysis of hundreds of un-electrified villages with different social, economical and geographical backgrounds. State-of-the-art planning models typically handle this computationally challenging task relying on highly-simplified technological characterizations, at the expense of a proper estimation of the cost-optimal potential of off-grid technologies, particularly micro-grids. In this paper, we propose a machine-learning method to improve such technological characterization while keeping the computational tractability of the problem under control. Firstly, field surveys from rural un-electrified villages in Bolivia are used as an input for a stochastic load generator model, creating several demand scenarios for a set of different village archetypes; secondly, renewable energy time series for representative locations of Bolivia are created using the NASA database. For each demand and renewables potential combination, a two-stage stochastic sizing model is adopted to obtain the corresponding cost-optimal micro-grid configuration. Finally, these data are used to train a Gaussian process regression with the levelized cost of energy (LCOE) as dependent variable and the daily average demand, renewable energy, and techno-economic characteristics of the components as independent variables. The results show that the trained algorithm is ultimately able to identify the LCOE of microgrids in given conditions, out of the training dataset, with satisfying accuracy and limited computational effort.@en