الفهرس 
Chapter 1Only 14 pages are availabe for public view
 |  | from | 115 |  |  |
| | | from | 115 | | |
Abstract
The current study presents a methodical framework to fully explore the possibility and potential of Artificial neural network (ANN) modeling to predict the performance of an indirect hybrid solar thermal storage system involving phase change materials (PCM) for domestic water heating application. A fully validated and tested numerical model simulating the performance of an indirect hybrid solar thermal storage system developed by others in previous work is adapted in the current study to produce the training, validation, and testing data necessary to adequately train and test the ANN model. To fully test the predictive and generalization ability of the ANN model; the studied system parameters (i.e., collector area, tank volume, demand temperature, PCM melting temperature, and PCM volume fraction) were varied over an extended range of values. To effectively train the ANN model; the training set must be optimized (i.e., Sampling method, and number of training samples). Three sampling techniques were investigated as follows:
1- Monte Carlo sampling method.
2- Latin Hypercube sampling.
3- Sobol Sequence sampling.
A python script is developed to run the validated numerical model against various input parameters using the three sampling methods to generate various training sets with various number of samples. An ANN model was developed for each training set and the testing results were plotted and compared. The performance of ANN models relies heavily on the internal model parameters known as hyperparameters. The current study investigates the effect of varying the model hyperparameters (i.e., learning algorithm, learning rate, activation function, number of hidden layers, and number of neurons per training layers) on the model predictive ability.
The results showed favorable results for Sobol sequence sampling compared to other sampling methods especially for lower number of training samples. It was observed that increasing the number of training samples had little to no improvement on the model testing score. Proper model hyperparameter tuning showed that a fully optimized ANN model can accurately mimic the performance of the studied system (i.e., coefficient of determination of 0.9999) while reducing the simulation time significantly (approximately five orders of magnitude).