The Goal
Forecasting electricity demand can help planners anticipate how much power needs to be generated. We used historical data from the Dominican Republic’s National Interconnected Electric System (SENI) to explore whether hourly demand data could support forecasts up to 30 days ahead.
What I Built
I worked across the technical process, from exploratory analysis and data preparation to modeling, evaluation, and preparing the model for deployment. I examined demand patterns and supporting variables, prepared the hourly data, and compared a 24-hour persistence baseline with Linear Regression, XGBoost, and an LSTM. The LSTM used the previous 24 hours of data to predict the next hour.
Process and Reflection
On the hourly test, the LSTM achieved a MAPE of 0.97%, compared with 1.20% for XGBoost and 1.29% for Linear Regression. Those scores measured short-horizon predictions. When we forecast further ahead, predictions stayed relatively close for the first few days, but errors grew substantially after roughly ten days. The project taught me that good next-hour accuracy does not establish reliability over a 30-day horizon. A future evaluation should measure performance across the full intended horizon.