Hybrid AI models enhance accuracy and sustainability of solar power forecasting

The study found that integrating spatial analysis from CNNs with temporal learning from LSTMs enabled the hybrid model to capture complex patterns in solar irradiance, temperature, and environmental variability. This capability is critical for forecasting in dynamic, high-irradiance regions such as those within the global solar belt.

Researchers have developed a hybrid AI model that significantly improves the accuracy and environmental sustainability of solar power forecasting. The study advances practical tools to support the transition of cities toward low-carbon energy systems, combining cutting-edge deep learning with carbon-conscious computing.

The study, “Enhancing Smart and Zero-Carbon Cities Through a Hybrid CNN-LSTM Algorithm for Sustainable AI-Driven Solar Power Forecasting (SAI-SPF),” was published in the Buildings journal. It evaluates multiple AI techniques for short-term solar power prediction, using real-world data from Benban Solar Park in Egypt and Sakaka Solar Power Plant in Saudi Arabia - two of the world’s largest utility-scale solar projects.

How can AI-driven forecasting optimize renewable energy for urban grids?

The researchers developed and compared seven forecasting models, including traditional machine learning, deep learning, and statistical methods. Among these, a hybrid model combining Convolutional Neural Networks (CNNs) with Long Short-Term Memory (LSTM) networks, known as CNN-LSTM, delivered the highest accuracy. The model achieved a Mean Absolute Percentage Error (MAPE) of 2.04% at Benban and 2.00% at Sakaka, far surpassing other techniques.

The study found that integrating spatial analysis from CNNs with temporal learning from LSTMs enabled the hybrid model to capture complex patterns in solar irradiance, temperature, and environmental variability. This capability is critical for forecasting in dynamic, high-irradiance regions such as those within the global solar belt.

Benban and Sakaka were chosen for their distinct yet complementary climatic conditions, Benban operating in a desert-arid environment and Sakaka in a semi-arid climate. The CNN-LSTM model’s consistent performance across both sites supports its adaptability to diverse geographic conditions within the high-solar-irradiance zone. The model’s accuracy was especially effective during midday peaks, when energy demand and generation typically surge.

Comparative analysis revealed that while Random Forest (RF) and Gradient Boosting Machines (GBM) models delivered respectable performance with moderate error rates and higher energy efficiency, they could not match the deep learning model’s predictive power. The ARIMA model, a traditional statistical tool, underperformed across all metrics.

Notably, the CNN-LSTM model was optimized using Bayesian hyperparameter tuning with a Tree-structured Parzen Estimator (TPE), ensuring minimal forecasting error and maximizing predictive reliability.

What role does environmental sustainability play in AI forecasting model design?

The study assesses carbon emissions generated during the training and inference phases of each model. Using the CodeCarbon tool, the researchers quantified CO₂ emissions in grams per computational hour for each algorithm, offering a rare evaluation of the environmental cost of AI in energy forecasting.

The CNN-LSTM model, though most accurate, also had the highest carbon footprint at 287 grams of CO₂ per hour. CNNs and LSTMs followed with emissions of 237 and 215 grams respectively. In contrast, traditional models such as ARIMA, RF, and SVM demonstrated markedly lower emissions - 75 grams, 102 grams, and 111 grams respectively.

This introduces a vital trade-off between performance and sustainability. While high-performing deep learning models like CNN-LSTM offer exceptional accuracy, their environmental cost must be carefully weighed, particularly in low-resource or edge-computing environments where energy efficiency is essential. For instance, in scenarios that demand frequent retraining or widespread deployment, lightweight models may offer better alignment with green AI objectives, even at a minor cost to forecasting precision.

To identify the key contributors to model performance, the researchers employed LIME (Local Interpretable Model-Agnostic Explanations). Solar irradiance and ambient temperature consistently ranked as the most influential features, contributing over 60% of the predictive power across both solar facilities. Wind speed and humidity also played meaningful roles, while static factors like panel orientation and month of the year were found to be less critical in short-term forecasting.

These insights highlight the importance of real-time, high-resolution meteorological data for accurate solar forecasting and reinforce the argument for integrating Internet of Things (IoT) sensors and weather monitoring systems into smart energy infrastructures.

How does the proposed framework support smart and zero-carbon city transitions?

The study introduces the Sustainable AI-Driven Solar Power Forecasting (SAI-SPF) framework, an integrated approach to optimizing solar energy production and grid management within urban environments. The framework incorporates real-time data acquisition, intelligent forecasting, greenhouse gas monitoring, and adaptive energy scheduling, forming a closed feedback loop that enhances both energy efficiency and carbon accountability.

Within this framework, one-hour-ahead forecasts are pivotal. They enable rapid decision-making for battery dispatch, load balancing, and curtailment minimization: essential functions for smart grid reliability. Moreover, carbon emissions estimates from model computation can inform environmentally responsible deployment strategies, such as selecting slightly less accurate models that significantly reduce CO₂ impact.

The model’s implementation on an edge-computing platform such as NVIDIA Jetson Nano, with an average inference latency of 210 milliseconds, demonstrates its viability for real-time applications. This makes it suitable for integration into decentralized systems like microgrids, where low latency and distributed intelligence are critical.

In addition to its technical contributions, the research outlines policy recommendations to align AI innovations with sustainability goals. These include mandating disclosure of model energy consumption and emissions, integrating real-time forecasting into energy contracts, and incentivizing energy-efficient algorithms through regulation. The authors advocate for embedding such models into national renewable energy strategies, particularly in emerging economies aiming for net-zero transitions.

By aligning with five core Sustainable Development Goals, clean energy (SDG 7), industry innovation (SDG 9), sustainable cities (SDG 11), climate action (SDG 13), and partnerships (SDG 17), the SAI-SPF system positions itself as a comprehensive solution for climate-resilient urban planning.

The study outlines future directions, including the integration of transformer-based models, domain adaptation techniques, and remote sensing data. It also highlights the need for lightweight AI algorithms for deployment in data-scarce or energy-constrained environments.