Abstract
In recent years, as photovoltaic (PV) power generation has rapidly increased on a global scale, there is a growing need for a highly accurate power generation forecasting model that is easy to implement for a wide range of electric utilities. Against this background, this study proposes a PV power forecasting model based on the generalized additive model (GAM) and compares its forecasting accuracy with four popular machine learning methods: k-nearest neighbor, artificial neural networks, support vector regression, and random forest. The empirical analysis provides an intuitive interpretation of the multidimensional smooth trends estimated by the GAM as tensor product splines and confirms the validity of the proposed modeling structure. The effectiveness of GAM is particularly evident in trend completion for missing data, where it is able to flexibly express the tangled trend structure inherent in time series data, and thus has an advantage not only in interpretability but also in improving forecast accuracy.
Highlights
This study focuses on publicly available weather forecast information and its application to PV forecasting methods for more general electric utilities, but even if we focus on such a target, many machine learning (ML)-based models have been proposed, such as artificial neural networks (ANN) [9] and support vector regression (SVR) [10], and ML has become a mainstream approach
It is emphasized that these points have often been overlooked in recent research. Motivated by these practical needs, this study proposes a forecasting model based on the generalized additive model (GAM) [21], a statistical approach, rather than ML
This study proposed and validated a GAM-based model with multidimensional tensor product splines to support the forecasting of PV power generation in practice
Summary
Publisher’s Note: MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations. Abdel-Nasser and Mahmoud [19] proposed a model using LSTM and compared it with multiple linear regression and NN, among others These previous studies have empirically shown that ML methods are superior in terms of forecast error reduction, but most forecasting methods using ML have challenges, such as high computational load and difficulty in interpretation, which is a well-known drawback in general. We verify the reliability of the models from the structural aspect by visualizing the smooth trends estimated using tensor product spline functions for the proposed GAM-based PV forecasting models and provide reasonable interpretations of the estimated trends. We conclude that the GAM-based model with multidimensional tensor product spline functions is superior in terms of interpretability, robustness, low computational load, and prediction accuracy.
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
