Advancing predictive modeling in conventional solar stills: a deep learning approach leveraging data augmentation and convolutional neural networks

[Display omitted] •Gaussian noise-based augmentation generated six synthetic samples per input, boosting data efficiency.•An optimized CNN-1D (5-128-128-128-1) was developed through systematic hyperparameter tuning.•Augmented CNN-1D achieved R2 = 0.97, RMSE = 0.04, MAE = 0.03 , outperforming baselin...

Celý popis

Uloženo v:
Podrobná bibliografie
Vydáno v:Energy conversion and management Ročník 346; s. 120565
Hlavní autoři: Migaybil, Hashim H., Gopaluni, Bhushan
Médium: Journal Article
Jazyk:angličtina
Vydáno: Elsevier Ltd 15.12.2025
Témata:
Convolutional neural networks (CNNs)
Data augmentation
Distillate water
Look-back window
Solar still
Support vector regression (SVR)
Data augmentation
Solar still
Support vector regression (SVR)
Convolutional neural networks (CNNs)
Look-back window
Distillate water
ISSN:0196-8904
On-line přístup:Získat plný text
Tagy: Přidat tag
Žádné tagy, Buďte první, kdo vytvoří štítek k tomuto záznamu!
Popis
Shrnutí:[Display omitted] •Gaussian noise-based augmentation generated six synthetic samples per input, boosting data efficiency.•An optimized CNN-1D (5-128-128-128-1) was developed through systematic hyperparameter tuning.•Augmented CNN-1D achieved R2 = 0.97, RMSE = 0.04, MAE = 0.03 , outperforming baseline CNN-1D and SVR.•Diagnostic tests confirmed homoscedastic residuals, minimal bias, and strong generalization.•Findings provide tools for predictive monitoring and optimization of solar desalination systems. Accurate forecasting of freshwater productivity from conventional single-slope solar stills is crucial for enhancing operational efficiency and minimizing capital costs. A persistent challenge in this domain is the scarcity of experimental data, which limits the training of reliable predictive models. This study proposes a data-efficient forecasting framework that integrates a one-dimensional convolutional neural network (CNN-1D) with time-series data augmentation. Gaussian noise sampled from N(0, 0.012) was applied exclusively to the training set, generating six augmented samples per instance. Both the augmentation factor (six) and the look-back window (seven days) were selected through systematic optimization, ensuring preservation of temporal dependencies. The CNN-1D architecture comprised three convolutional layers with 128 filters, ReLU activations, a flattening stage, and a dense regression output layer. Hyperparameters—including learning rate, batch size, kernel size, and regularization strength—were fine-tuned using Tree-structured Parzen Estimator (TPE) optimization with a maximum of 50 trials, where the best-performing configuration achieved the lowest loss. Model training employed a feed-forward backpropagation algorithm with 365 daily observations to predict freshwater yield (Pstd, L/day). Benchmarking against an optimized support vector regression (SVR) model with a radial basis function kernel revealed that the augmented CNN-1D achieved superior performance (RMSE = 0.04, MAE = 0.03, OIMP = 0.97), consistently outperforming both the baseline CNN-1D and the optimized SVR. Residual analyses confirmed its robustness, minimal bias, and strong generalization across unseen data. These findings demonstrate that combining augmentation with hierarchical feature extraction enables a scalable and computationally efficient predictive tool for solar still performance, offering significant potential for sustainable freshwater management in arid and data-constrained regions.
ISSN:0196-8904
DOI:10.1016/j.enconman.2025.120565