A Neural Network-Based Causal Model for Electricity Demand Estimation in Remote Areas: A Case Study in El Espino, Bolivia

Abstract

Designing microgrids in remote areas is challenging due to the lack of reliable and high-quality electricity demand data. These limitations arise from technological, economic, and logistical constraints, making it difficult to estimate demand—especially when daily demand curves cannot be fully constructed due to missing or incomplete data. Traditional methods, which rely on consistent and comprehensive datasets, often prove ineffective in such conditions. To address this issue, this paper introduces a novel causal modelling approach, implemented using a neural network, to uncover the underlying relationships between key influencing factors—such as temperature, humidity, time of day, and seasonal variations—and electricity demand. Rather than requiring complete hourly demand curves as inputs, the proposed approach leverages available data to infer demand patterns more effectively. We propose a neural network architecture that aims to capture causal dependencies in electricity demand by encoding input features into a high-dimensional latent space. Using an encoderdecoder structure, the encoder maps inputs to a latent space designed to preserve potential causal relationships, while the decoder generates the demand estimation. This approach hypothesizes that this configuration may help to get causal dependencies. To evaluate this, we compared our model against a simpler neural network architecture characterised by a triangular layer structure. Using real-world data from El Espino, Bolivia, our model achieved a Mean Squared Error (MSE) of 0.0511 with the Adam optimiser, representing a 61.8% improvement over the simpler neural network architecture. A sensitivity analysis further confirmed the relevance of selected input variables, showing that excluding temporal-based features, such as the month of the year and weekend indicator, increased estimation error, with an 11.7% increase in MSE. These findings highlight the model’s effectiveness in handling data limitations and its potential as a scalable solution for electricity demand estimation in remote areas.