Mimicking chimpanzee hunting behavior to improve PV prediction models – pv magazine International

11 September 2024 Lior Kahana

A scientific group led by researchers from the German-Jordanian University has analyzed the effect of the so-called Chimp optimization algorithm (ChOA) on various machine learning (ML) models for predicting PV yield production.

The ChOA is based on the cooperative hunting behavior of chimpanzees in the wild, mimicking the way they work together to take down prey, as is common among small mammals. They usually operate in a group of three or four hunters, initially herding and blocking the prey, then chasing and attacking it.

The algorithm explores different parameter combinations to achieve the most promising result. It was used by the scientists to optimize the hyperparameters for five types of ML models. These include multiple linear regression (MLR), decision tree regression (DTR), random forest regression (RFR), support vector regression (SVR), and multi-layer perceptron (MLP).

“The effectiveness of this contribution is verified using data from a real case study, using several performance metrics from the literature, including root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R2),” the researchers explained.

Hyperparameters are external configurations that are set before the learning process begins, control the learning process, and do not change during training. Hyperparameters – such as the learning rate in neural networks – influence the training dynamics and can therefore significantly affect the effectiveness of models.

All five models, with and without ChOA, were trained on 948 datasets and tested on 362 datasets. The datasets were collected between 2015 and 2018 from a 264 kW PV system installed on a rooftop of the Applied Science University in Amman, the capital of Jordan. The tilt angle of the system was set to 11 degrees and the azimuth angle to -36 degrees. Meteorological variables such as wind speed, relative humidity, ambient temperature and solar radiation were measured from a nearby weather station.

“Amman, Jordan, has a Mediterranean climate characterized by hot, dry summers and cool, wet winters,” the researchers added. “The average annual temperature is 17.63 °C and the mean annual global horizontal radiation is 2040.2 kWh/m2.”

Through this analysis, the scientists found that all models improved performance by fine-tuning the hyperparameters using ChOA.

“DTR showed significant improvements, with the test RMSE decreasing to 1.972 and R2 increasing to 0.951,” they explained. “The RFR model showed notable improvements, with RMSE values ​​decreasing to 1.773 for training and 1.837 for testing, and R2 values ​​increasing to 0.964 for training and 0.963 for testing. The SVR model experienced the most notable improvement, with the test RMSE decreasing to 0.818 and R2 increasing to 0.977.”

After ChOA optimization, MLP showed the best results in predicting PV power yield. In particular, it could achieve 0.503, 0.397, and 0.99 in RMSE, MAE, and R2, respectively. “The ChOA effectively fine-tuned the parameters, resulting in improved model fitting, less overfitting, and improved generalization compared to two other widely used optimization algorithms from the literature: particle swarm optimization (PSO) and genetic algorithm (GA),” the team concluded.

The results were presented in “Improving solar photovoltaic energy production prediction using different machine learning models tuned to the Chimp optimization algorithm,” published in Scientific reportsThe group included scientists from the German-Jordanian University, the University of Jordan, Al-Balqa Applied University and Tuskegee University in Alabama.