Deep Learning-Driven Irradiance Forecasting for Adaptive Solar Farm Output Maximization

Kumar, Hemant; Sokhal, Gurpreet Singh; Singh, Rupinder; Debnath, Suman

Volume 2, Issue 2 (11-2026) AISEE 2026, 2(2): 115-129 | Back to browse issues page

Mendeley

Zotero

RefWorks

Kumar H, Sokhal G S, Singh R, Debnath S. Deep Learning-Driven Irradiance Forecasting for Adaptive Solar Farm Output Maximization. , AISEE. 2026; 2 (2) :115-129
URL: http://aisesjournal.com/article-1-53-en.html

Deep Learning-Driven Irradiance Forecasting for Adaptive Solar Farm Output Maximization

Hemant Kumar ^*

, Gurpreet Singh Sokhal

, Rupinder Singh

, Suman Debnath

Punjabi University, Patiala, Punjab, India

Abstract: (8 Views)

The intermittent nature of solar radiation causes uncertainty in solar PV power system operations, posing challenges to system stability and resulting in higher operational costs. The present study aims to tackle this significant gap between irradiance prediction and maximum power point tracking (MPPT) optimization by introducing a hybrid deep learning framework. The novel Spatiotemporal Deep Irradiance Forecasting Engine (SDIFE) is a hybrid CNN-LSTM-Attention architecture with quantified uncertainty bounds for the generation of probabilistic 15-minute ahead global horizontal irradiation (GHI) forecasts. The forecasts directly feed into a Forecast-Informed Adaptive MPPT Controller (FI-AMC) that changes voltage setpoints of the inverters ahead of time based on the forecasted trajectory of the irradiance. The proposed control structure was validated by applying it to a 50 MW solar PV utility-scale plant for 24 months of operational data, which resulted in a normalized root mean square error (NRMSE) of 3.12% for the GHI forecasting accuracy (a 34.7% improvement over the standalone LSTM benchmark) and a 99.4% MPPT tracking efficiency during rapid cloud transients, equivalent to a 4.8% annual energy yield improvement over conventional perturb-and-observe control. The results of ablation studies support these findings and indicate that each architectural innovation makes measurable contributions to performance and that the trust adaptive forecast-measurement fusion parameter is most important. A computational latency of 2.3s means the edge deployment can be done within the 5-minute control interval, and preliminary techno-economic analysis suggests a payback period of 2.3 years for installations in a utility fleet. The results show that tightly coupled forecasting and control architectures are significantly more effective than optimization approaches that are decoupled, providing a route towards anticipatory, and not reactive, solar farm management.

Keywords: Solar irradiance forecasting Deep learning Maximum power point tracking Adaptive control, Physics-informed neural network Convolutional neural network

Full-Text [PDF 1015 kb] (3 Downloads)

Type of Study: Research | Subject: Special
Received: 2026/06/22 | Accepted: 2026/06/27 | Published: 2026/11/11

* Corresponding Author Address: Department of Mechanical Engineering, Punjabi University, Patiala, Punjab, India