Poster_Ibtihal_Ait_Abdelmoula

Ibtihal Ait Abdelmoulaa,b* , Said Elhamaouib, Omaima Elalanib, Abdellatif Ghenniouib , Mohamed El Aroussia a SIRC/LAGeS-EHTP, Hassania School of Public Works, Casablanca, Morocco b Green Energy Park - UM6P, Benguerir, Morocco * Corresponding author. Tel.: +212 676170872 e-mail address: aitabdelmoula@greenenergypark.ma Recording data measurements Resampling and Synchronisation Removing outliers, gaps and duplicates Replacing negative values with zeros Remove days with very few records Data cleaned and ready for analysis 1. MOTIVATION KEY FINDINGS • Investigation of a model built employing stacked Machine Learning Algorithms (RF, XGBoost, MLR), historical electrical and meteorological data, and feature engineering of input parameters to enhance the model accuracy. The stacked model gave better results than individual models scoring a RMSE of 386.31 W for system 1 and 131.54 W for system 2. • Validation of the model on diverse PV technologies, including differences in module type, plant size and exposure duration, with the aim of examining their impact on the PV power prediction. The Moroccan Solar Plan is considered one of the most ambitious renewable programmes in the African continent. Characterized by its high insolation and arid/semi-arid climate, Morocco relies on R&D to efficiently integrate renewables into its national grid. Understanding the distinctive features of each environment as well as how they affect photovoltaic power is the first step towards developing intelligent and data-driven maintenance algorithms suited to the region s environmental context. A Stacked Machine Learning Model is used in this study to make hourly power predictions of two PV systems varying in size, age and technology. MATERIALS AND METHODS The case study was conducted within the Green Energy Park (GEP) research and test platform. The studied PV systems have a total capacity of 16.56 kWp (System 1) and 5.34 kWp (System 2). The electrical and meteorological parameters are monitored with a 2-min resolution using adapted dataloggers. The dataset used in the study ranges from October 2021 to March 2022 and represents the winter and autumn seasons known for the high fluctuation in solar irradiance and prediction complexity. 3000 HOURS OF SUNSHINE PER YEAR AND GHI RANGING FROM 1,753 UP TO 2,264 KWH/M² IN THE DESERT MOROCCO TARGETS 52 % RENEWABLES IN THE MIX BY 2030 (AROUND 6 GW OF SOLAR CAPACITY) Copyright : Iresen/ Mines ParisTech/ Transvalor/ Armines MOTIVATION Figure 1 : Solar radiation map of Morocco Figure 2 : Representation of the testbed used in the case study Figure 3 : Data Handling methodology RESULTS ONGOING WORK Figure 4 : Architecture of the stacked model proposed in the study MODEL IMPLEMENTATION FEATURE COMBINATION 1 (BASIC) : GHI, GTI, Ambient Temperature, Humidity, Wind speed, Module temperature, FEATURE COMBINATION 2 (ENGINEERED) : COMB 1 + Hour, Prev_Hour1, Prev_Hour2, MA_Power, Prev_Day MO DE L INPU TS • Testing the model on a larger dataset to study the trade-off between the size of the dataset and the prediction performance on both systems. • Evaluating the complexity and implementation feasibility of the models on an edge computing environment. A Photovoltaic Power Prediction Approach Enhanced by Feature Engineering and Stacked Machine Learning Model Figure 5 : Performance Comparison for the studied systems – For System 1 (a), the best results were found using the engineered features whereas for System 2 (b) the basic features gave better results. The plots show only the best-case scenario for each system. The scatter plots of measured vs predicted power for each model are shown in figures (c) and (d) (a) (b) ACCESS FULL PAPER METEO STATION PV SYSTEM 1 PV SYSTEM 2 (c) (d)SCAN FOR MORE INTERACTIVE MAPS !