This thesis presents a multi-objective optimization of a centrifugal pump at multiple operating points (MOMP). The pump driven by a BLDC motor was designed for a hot water circulation system. The primary motivation was to minimize the energy consumption during transient operations according to the extended product approach and the Energy-Efficiency-Index of the recent EU regulation. We set the optimization objectives based on respective weighted sums of the heads and efficiencies at four operating conditions of the standardized flow-time profile. The main design parameters were determined by sensitivity analysis using the Elementary Effects method. The design of experiment was conducted using Latin Hypercube Sampling method. The Response Surface Methodology was adopted to construct the surrogate models. A multi-objective optimization was conducted by using a genetic algorithm. The optimization of the impeller following this procedure produced a Pareto front. The volute was optimized on a similar procedure that focused on reducing the sidewall circulating flows. The final design was far superior to the original one in terms of the efficiency and power. MOMP outperformed those using the multi-objective optimization at the best efficiency point (BEP) and the single-objective optimization at BEP throughout the objective space although all the methods performed similarly in a local region. Therefore, the design of a pump with variable flow will benefit from this MOMP approach.
The second objective of this work involves investigating the application of Long Short-Term Memory (LSTM) networks, a specialized form of recurrent neural network, for the prediction of domestic hot water demand. Given the temporal dependencies inherent in hot water consumption patterns, LSTM networks with their ability to learn long-term dependencies in sequence data present a promising solution. The analysis incorporates the design and implementation of both single-layer and multiple-layer LSTM models, aiming to understand their capacity to handle complex, nonlinear relationships in time-series data. The single-layer LSTM model encounters some challenges in predicting unexpected peaks in demand, particularly on weekends and holidays. In contrast, the multiple-layer LSTM model demonstrates superior performance, providing more accurate predictions and demonstrating lower error rates. The model's efficiency is gauged using various statistical metrics such as the root mean square error (RMSE) and the mean absolute scaled error (MASE), confirming the improved precision of the multi-layer LSTM model. The findings underscore the potential of LSTM networks as a robust tool for managing and controlling hot water resources effectively.