Soiling is a major obstacle to exploiting the full potential of concentrated solar power (CSP) systems, particularly in semi-arid regions where dust accumulation is high. This study introduces a machine learning methodology to predict soiling, with the aim of improving cleaning programs, enhancing plant performance and reducing operating costs. The methodology incorporates a comprehensive dataset of meteorological variables (temperature, humidity, wind speed, precipitation, and Aerosol Optical Depth) collected at the Green Energy Park in BenGuerir, Morocco, along with 16 months of continuous soiling measurements. Three distinct deep learning architectures, including an autonomous Long Short-Term Memory (LSTM), a hybrid CNN-LSTM and a CNNBiLSTM model with attention mechanism, are developed and evaluated using a recursive prediction strategy for multi-step forecasting. Results demonstrate that the CNN-LSTM model achieves the highest predictive accuracy with an R2 of 0.88, MAE of 0.03, and RMSE of 0.03, outperforming the stand-alone LSTM model (R2 = 0.73, RMSE = 0.05) and showing comparable performance to the CNN-BiLSTM-attention model (R2 = 0.87, RMSE = 0.04). The model successfully captures both daily fluctuations and seasonal soiling trends, with maximum deviation from measured values of 0.0986 and mean deviation of 0.0219. These improved predictions enable more strategic cleaning protocols, allowing operators to optimize maintenance schedules based on actual soiling forecasts rather than fixed intervals, thereby maximizing mirror reflectance while minimizing water consumption, labor costs, and operational expenditures in semi-arid CSP installations.