Search

Browse

• Communities
• Authors
• Year
• Titles
• Subjects

My Space

• Login

Artificial neural network analysis of a heat pump system and its efficiency inside a broiler house

Chachalis Alexandros

Full record

---

Summary

In the context of this thesis, the utilization and training of artificial neural networks is examined to simulate the coefficient of performance (COP) of a heat pump (HP) system inside a broiler house. The heat pump system aims to cover the thermal needs of the broiler house that is located in Kavala and is housing a total of 10000 broilers.Artificial neural networks simulate the coefficient of performance of the heat pump being utilized for heating, cooling and dehumidifying each time. As a first step, the dataset being used to design the ANN is provided by Tyris et.al ,2023, the research team that designed and studied the heat pump system, and includes values of indoor temperature and relative humidity, broilers’ thermal loads as well as the simulated values of COP that their dynamic model calculated. Initially, the input table and the target vector which the ANN utilizes for training, were created. The input table consists of the indoor temperature and relative humidity data within the facility as well as the heat load of the animals. In addition, for faster convergence of the ANN, preprocessing was performed by normalizing the input data using the Z-score standardization method. The target vector, contains the simulated values of the coefficient of performance estimated by the dynamic model of Tyris et.al ,2023. The above data on which the ANN are trained cover the periods (35 days duration) of winter (1/1- 5/2) and summer (31/5-5/7) and include 54000 values for each period as the time step between data is 1 minute.After completing the preprocessing of the input data, the training of the ANN was performed using Matlab's Neural Fitting Tool (nftool) and the training algorithm used was Levenberg-Marquardt (LMA). Training of the ANN took place for different parameters in each test/simulation by increasing the number of hidden nodes or changing the training percentages.The objective of neural network training is to find the model and its parameters that give the optimal results with the selection criteria of these being the root mean square error (RMSE) and the correlation coefficient (R). More specifically, it is desirable that RMSE is the minimum and R is the maximum among the selected hidden nodes and training percentages.In this thesis, the ANN are trained for the operation of the heat pump system during the winter and summer periods when heating and cooling are used respectively, while in both periods dehumidification is also used.Overall, the LM algorithm achieved minimum errors and maximum correlation coefficient of 10-2 and 0.999, respectively, for all periods and thermal modes. Despite these very satisfactory, results, it was observed that the neural network cannot estimate the outliers (values) for some hidden node trials, possibly because of the lack of intermittent values of the coefficient of performance (COP) based on which the neural networks were trained.

Available Files

Services