A Hybridized CNN-LSTM-MLP-KNN Model for Short-Term Solar Irradiance Forecasting

Abstract

The accurate prediction of solar irradiance is essential for maximizing renewable energy production. Despite some researchers struggling to achieve sufficient prediction accuracy and model quality, we are working on developing a forecasting model for solar radiation using advanced artificial intelligence techniques. Solar energy offers a sustainable alternative to fossil fuels and has a wide range of applications, making the development of an effective forecasting model crucial. This paper proposes a hybrid model that combines deep learning and machine learning methods, denoted as CNN-LSTM-MLP-KNN. By combining Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), Multilayer Perceptron (MLP) and K-Nearest Neighbors (KNN), we aim to enhance the accuracy and effectiveness of time series forecasting models. The study focuses on extracting spatial and temporal patterns from sun irradiance data using CNN and LSTM and uses MLP to examine intricate connections. The KNN regressor algorithm is employed to make non-parametric forecasts based on the nearest neighbors, resulting in the final forecasting for solar irradiance. Our work relies on the historical Karachi dataset from 2017, 2018, and 2019, sourced from the NSRDB, which provides sun irradiance measurements at a 15-minute time interval using accurate meteorological instrumentation. This dataset covers a 3-year period and comprises 105120 samples with 24 features. Our proposed model provides more accurate predictions compared to the most recent published models, with an R2 score of 0.9874, MSE of 0.0009, RMSE of 0.0311, and MAE of 0.0118. The source code is publicly accessible at https://github.com/Short-Term-Solar-Irradiance-Forecasting.