Comparative analysis of Elman and Jordan Recurrent Neural Networks for Solar Radiation and Air Temperature Prediction Using Backpropagation Variants

Abstract

This study compares the effectiveness of Elman Recurrent Neural Networks (ERNN) and Jordan Recurrent Neural Networks (JRNN) for predicting solar radiation and ambient temperature in Ouarzazate, Morocco. Data collected at 10-minute intervals from three meteorological stations (OUA_001, OUA_002, and OUA_003) over a 3–4-year period (2018-2022) were analyzed. The dataset was split into training (75%) and validation (25%) subsets to develop and test the models. The research systematically explored different network architectures with 1-2 hidden layers containing 4-12 neurons, applying three learning algorithms: BackPropagation (BP), BackPropagation with Momentum (BPM), and Resilient Backpropagation (Rprop). Key hyperparameters were optimized within specific ranges, including learning rates (0.00001-0.4) for BP/BPM and weight decay exponents (0.00001-4) for Rprop. Input variables included date, temperature, solar radiation, wind speed, relative humidity, precipitation, and atmospheric pressure in various combinations. Performance evaluation using Nash-Sutcliffe Efficiency (NSE) and Index of Agreement (d) revealed high prediction accuracy for both model types, with values exceeding 0.9 during validation. JRNN with BPM performed best at station OUA_001 (NSE: 0.909 for radiation, 0.971 for temperature), while ERNN with BPM demonstrated superior performance at station OUA_002 (NSE: 0.978 for radiation, 0.976 for temperature). At station OUA_003, both models showed comparable results when using BP. Despite the high overall accuracy, both models exhibited limitations in predicting extreme solar radiation values, particularly during nighttime hours. The study concludes that ERNN and JRNN are effective tools for short-term prediction of solar radiation and temperature in arid regions like Ouarzazate, though further refinement is needed to better capture extreme values and improve prediction accuracy during transitional periods.