Performance Prediction and Optimization of Solar Water Heater via a Knowledge-Based Machine Learning Method

Introduction

Predicting the thermal performance of solar energy systems is of huge challenge due to the complexity of the internal structures. In fact, the measurement of thermal performances of a typical solar energy system (e.g., solar water heater (SWH)) requires a lot of time, economic costs and labors. Due to the complicated structures, conventional physical and mathematical models usually fail to estimate their thermal performances. These problems not only dramatically hinder the acquisition of the thermal performances of solar energy systems, but also block the possibility of optimizing their thermal performance.

In the past decades, scientists have come up with a powerful prediction method to address these problems. People found that a knowledge-based machine learning model can help precisely predict the performances of some energy systems utilizing some simple independent variables as the computational inputs. With a proper machine learning algorithm, people only need to acquire a sufficient experimental database as well as perform the model training and testing, and then a predictive model can be acquired. During the training process, machine learning can “learn” the non-linear relationship between the independent and dependent variables via a “black box” fitting, and subsequently perform the predictions. The research group of Prof. S. Kalogirou, from Cyprus University of Technology, has conducted a majority of the pioneer application research on the prediction of thermal performances for energy systems (Kalogirou, 1999), leading to huge positive engineering impacts during the last two decades. Subsequently, relevant studies have become increasingly popular all over the world. Meanwhile, there are more and more new or revised machine learning algorithms developed. Among various machine learning approaches, there are some most widely used algorithms, such as artificial neural network (ANN) (Kalogirou, 1999), support vector machine (SVM) (Suykens & Vandewalle, 1999) and extreme learning machine (ELM) (Huang et al., 2006). ANN is the most prevalent algorithm due to its long history and powerful predictive capacity. The general schematic structure of an ANN model is presented in Figure 1.

Figure 1.

General schematic structure of a typical ANN model. The empty circles represent the neurons. All the neurons interconnect with other neurons in the adjacent layer(s). Each neuron in the input layer represents an independent variable. The neuron in the output layer represents the dependent variable.

Reproduced with permission from Reference (Li, Liu, Liu, & Zhang, 2017).

So far, despite the great progress of machine learning, engineering and industrial requirements have been rising in recent years: how to cost-effectively design and optimize a solar energy system by utilizing machine learning? Now, machine learning is a proven powerful tool for varieties of numerical predictions, and people are trying to make full use of its predictive power, as well as provide a good optimization strategy in order to acquire higher performances. Nevertheless, to the best of the authors’ knowledge, very few research reports have mainly focused on the relevant studies (Peng & Ling, 2008). Recently, the authors have found that with a sufficiently large experimental database, the machine learning models are not only able to give excellent predictive performance for SWHs, but also assist an efficient and promising optimization of the thermal performances of SWHs, with the usage of a high-throughput screening (HTS) process (Zhijian Liu, Li, Liu, Yu, & Cheng, 2017). This is, so far, the first study of HTS on the optimization of energy system.