Year

2025

Season

Fall

Paper Type

Master's Thesis

College

College of Computing, Engineering & Construction

Degree Name

Master of Science in Electrical Engineering (MSEE)

Department

Engineering

NACO controlled Corporate Body

University of North Florida. School of Engineering

Committee Chairperson

Dr. Touria El Mezyani

Second Advisor

Dr. Xudong Liu

Third Advisor

Dr. Hemani Kaushal

Department Chair

Dr. Alan Harris

College Dean

Dr. William Klostermeyer

Abstract

Accurate short-term forecasting of solar power generation is critical for the reliable and cost-effective operation of renewable-based microgrids, where sudden weather-induced variability can compromise grid stability, battery scheduling, and energy trading decisions. Traditional physical and statistical models struggle to capture the complex non-linear relationships and localized weather effects, while individual deep learning architectures often exhibit systematic biases such as chronic under-prediction of peak generation. This thesis proposes a novel Cross-Feedback Ensemble framework that combines the complementary strengths of Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and one-dimensional Convolutional Neural Network (1D-CNN) models through an iterative cross-feedback mechanism and a linear meta-learner. Unlike conventional stacking or error-gated ensembles, the proposed architecture allows each base model to receive the real-time predictions of the others as additional input features during retraining phases, enabling mutual error correction and richer representation of temporal, short-term, and spatial patterns in weather-to-power relationships. The models were developed and rigorously evaluated using four years (2016–2019) of hourly solar generation and meteorological data for Germany from the Open Power Systems Data (OPSD) platform. The proposed Cross-Feedback Ensemble achieved a Mean Absolute Error (MAE) of 512.3 MW, Mean Squared Error (MSE) of 883,954.54 MW², and coefficient of determination (R²) of 0.9756 on the hold-out test set — representing improvements of approximately 11.9%, 14%, and 0.41%, respectively, over a baseline stacking ensemble and even larger gains compared to standalone LSTM, GRU, and CNN models. Most notably, the cross-feedback mechanism virtually eliminated the systematic under-prediction bias observed in individual deep learning models. The results demonstrate that explicit inter-model feedback and dynamic error correction significantly enhance forecasting accuracy and robustness in real-world variable weather conditions. The proposed framework offers a practical, high-performance solution for microgrid energy management systems and can be extended to multi-source renewable forecasting and real-time operational deployment.

Download

Share

COinS