Coati-Inspired Algorithm Boosts Solar Radiation Forecast Accuracy in Indian Study

Solar power lives and dies by the forecast. Grid operators need to know, hour by hour, how much sunlight will strike their panels, and even small prediction errors ripple through scheduling, storage dispatch and electricity trading.

The work, published in Environmental Systems Research, combines two ideas that machine learning researchers have pursued separately for years: careful feature selection and metaheuristic optimization of network weights. Led by T. K. Nida Fariz and S. Sharief Basha, the team built a model they call COA-ANN, in which the coati algorithm searches for the best set of connection weights and biases for a feedforward neural network, replacing the gradient descent procedure that conventional networks use during training.

The motivation is rooted in a well-known weakness of standard neural network training. Training an ANN typically means minimizing an error function, such as mean squared error, by iteratively adjusting connection weights with gradient descent. But gradient-based methods can become trapped in local minima, points in the error landscape that look optimal from nearby but are far from the best possible solution. Metaheuristic algorithms, which balance exploration of new regions of the search space with exploitation of promising ones, offer a way to escape those traps. Previous studies have shown that metaheuristic-trained networks, including those tuned by particle swarm optimization and the Grey Wolf Optimizer, can outperform traditional backpropagation.

The Coati Optimization Algorithm, developed by Mohammad Dehghani and colleagues in 2023, mimics two foraging and survival behaviors of coatis, members of the raccoon family. In the first phase of the algorithm, modeled on coatis hunting iguanas, part of the population climbs toward the best solution found so far, representing exploration, while the rest search the ground after the prey metaphorically falls, a random repositioning step. In the second phase, modeled on coatis fleeing predators, each candidate solution probes random nearby positions within shrinking local bounds, refining the search. Candidate solutions that improve the objective function are kept; others are discarded. In the hybrid model, each individual coati encodes a complete vector of the neural network’s weights and biases, and the fitness function is simply the mean squared error of the network on the training data.

The data behind the study came from a pyranometer at the School of Mechanical Engineering at VIT Vellore, capturing hourly solar radiation along with eighteen weather variables for March 2023. Before modeling, the researchers filtered the dataset, excluding nighttime records and periods with very low solar elevation angles, below 10 degrees, since solar radiation is physically absent at night and such records would push the model toward trivial zero-radiation predictions. Input variables were scaled to the range zero to one using min-max normalization, a linear transformation that helps activation functions operate in their effective range during training.

Feature selection played a central role. The team computed Pearson correlation coefficients between each weather variable and solar radiation, keeping only variables whose absolute correlation exceeded a threshold of 0.1. This reduced dimensionality by discarding irrelevant or redundant inputs and, the authors argue, improved both the efficiency and the accuracy of the downstream model. The selected inputs were outdoor temperature, dew point, wind speed, humidity and the ultraviolet index. The final network architecture was compact: five input variables feeding a single hidden layer of nine nodes with sigmoid activation, producing one output, a configuration denoted 5-9-1. Eighty percent of the data was used for training and the remaining twenty percent reserved for testing.

The benchmarking results were one-sided. The COA-ANN model achieved a root mean square error of 35.6009, a mean absolute error of 2.6557, and a mean absolute percentage error of 4.558 percent, the lowest error metrics among all tested approaches, along with the highest coefficient of determination at 97.244 percent. The comparisons were run against a standard ANN trained by gradient descent, a PSO-ANN hybrid and a GWO-ANN hybrid, using identical error rubrics. The mean absolute percentage error rose to 5.992 percent for GWO-ANN, 7.124 percent for PSO-ANN and 9.794 percent for the backpropagation network. The gradient descent baseline used a single hidden layer of ten units and a learning rate of 0.01, yet still trailed the metaheuristic hybrids.

The authors also placed their results alongside previously published solar radiation forecasting methods. In that comparison, drawn from earlier literature, the average RMSE was 59.2855 for a biogeography-based optimization hybrid ANN, 84.1469 for a recurrent neural network approach and 135.8127 for a support vector regression model, all higher than the 35.6009 reported for COA-ANN on this dataset. Scatter plots and regression analyses of predicted versus observed radiation showed the closest agreement for the coati-optimized model, and convergence plots suggested the algorithm settled on stable weight configurations during training.

Computational efficiency was a further selling point. Running in MATLAB R2023a on a laptop with 16 gigabytes of RAM, the model produced a forecast for an entire month in under 1.2 seconds on average, a speed the authors attribute to the compact network and the efficiency of the optimization loop. They argue this combination of accuracy and speed makes the approach practical for operational settings, where forecasts must be regenerated frequently as new weather observations arrive. The study’s scope, however, is deliberately narrow: the data come from a single district during a single month, and the authors note that the model was designed to predict solar radiation for that specific geographic area under the meteorological conditions represented in the training data. Generalization to other climates, seasons or years remains to be demonstrated, and the authors themselves flag limited generalizability to fresh data as a persistent challenge in the solar forecasting field.

The broader context gives the work its urgency. India’s electricity system remains heavily dependent on fossil fuels, and the paper cites national estimates that the country could theoretically generate more than 5,000 trillion kilowatt-hours of solar energy annually, with the potential to install 748 gigawatts of solar capacity on just three percent of its wasteland. Hourly radiation forecasts feed directly into the practical side of that buildout, supporting generation estimation, grid integration, storage sizing and real-time adjustments for tracking solar installations. The authors suggest future work could extend the framework with ensemble structures or additional input variables. For now, the study stands as a controlled, single-site demonstration that a 2023-vintage metaheuristic, borrowed from the foraging habits of a tropical mammal, can squeeze measurably more accuracy out of a small neural network than the standard training recipe, at least for one month of Tamil Nadu sunshine.

Subject of Research: Hybrid artificial neural network with coati optimization and correlation-based feature selection for hourly solar radiation forecasting in Vellore, India

Article Title: Enhancing solar radiation forecasting via feature selection in hybrid artificial neural networks: a comparative study with coati optimization

Article References: Fariz, T. K. N., & Basha, S. S. (2026). Enhancing solar radiation forecasting via feature selection in hybrid artificial neural networks: a comparative study with coati optimization. Environmental Systems Research, 15(1), Article 44. https://doi.org/10.1186/s40068-026-00490-2

Image Credits: AI Generated

DOI: 10.1186/s40068-026-00490-2

Keywords: solar radiation forecasting, artificial neural networks, Coati Optimization Algorithm, metaheuristic optimization, feature selection, Pearson correlation, renewable energy, India, Vellore, hybrid machine learning, photovoltaic energy, time series prediction