Forecasting and Probabilistic Scenario Generation of Load and Meteorological Factors using Artificial Intelligence

Forecasting of electrical load and meteorological factors is essential for energy planning of modern power systems incorporating renewable energy sources. However, the loss of variance in the time series is a problem when using recursive strategies in short-term forecasting and generating future simulations with horizons of one year and longer. This study proposes a comprehensive simulation framework with recursive one-step-ahead forecasting model targeting one-year scenario generation for four different parameters (electrical load, wind speed, temperature, and solar radiation). In the application of the methodology, the suitability of Support Vector Regression (SVR), Random Forest (RF), Extreme Gradient Boosting (XGBoost), and Long Short-Term Memory (LSTM) models was evaluated, and in the evaluation performed with error metrics on the test data, LSTM demonstrated the highest accuracy. Subsequently, the proposed methodology was implemented using the LSTM model. A cascade structure was then developed, using both historical load data and generated future data for population and temperature as inputs. During a one-year simulation employing a proposed recursive strategy, time-stratified residual bootstrapping was integrated to prevent variance dampening, successfully preserving the statistical nature of the historical data.