KNN and Time Series Based Prediction of Power Generation from Renewable Resources
Ismum Ul Hossain, Mohammad Nahidul Islam
TL;DR
This study addresses the challenge of forecasting renewable power generation by comparing two time-series approaches, KNN regression and SARIMA, using a ~30-year dataset across solar, wind, and hydro sources. The methodology includes careful dataset preparation, parameter tuning (SARIMA with $P$, $D$, $Q$, and $s$ optimized via AIC/BIC; KNN with cross-validation for choosing $K$), and 10-year ahead forecasting to evaluate predictive performance. Results show that KNN and SARIMA offer complementary strengths: SARIMA captures seasonal and temporal structure, while KNN provides accurate local forecasts for individual sources. The findings have practical implications for grid reliability, energy trading, and policy planning, and suggest that hybrid models could further enhance forecast accuracy.
Abstract
As the world shifts towards utilizing natural resources for electricity generation, there is need to enhance forecasting systems to guarantee a stable electricity provision and to incorporate the generated power into the network systems. This work provides a machine learning environment for renewable energy forecasting that prevents the flaws which are usually experienced in the actual process; intermittency, nonlinearity and intricacy in nature which is difficult to grasp by ordinary existing forecasting procedures. Leveraging a comprehensive approximately 30-year dataset encompassing multiple renewable energy sources, our research evaluates two distinct approaches: K-Nearest Neighbors (KNN) model and Non-Linear Autoregressive distributed called with Seasonal Autoregressive Integrated Moving Average (SARIMA) model to forecast total power generation using the solar, wind, and hydroelectric resources. The framework uses high temporal resolution and multiple parameters of the environment to improve the predictions. The fact that both the models in terms of error metrics were equally significant and had some unique tendencies at certain circumstances. The long history allows for better model calibration of temporal fluctuations and seasonal and climatic effects on power generation. The reliability enhancement in the prediction function, which benefits from 30 years of data, has value to grid operators, energy traders, and those establishing renewable energy policies and standards concerning reliability