1. Introduction
In the last decades, climate change, global economic instability, and increasing dependence on fossil fuels have become urgent challenges requiring an accelerated transition to more sustainable and efficient energy sources [
1]. In this context, renewable energy, including geothermal energy, emerges as a promising alternative to provide a clean and constant energy supply [
2]. Unlike solar and wind energy, whose technologies have evolved significantly due to their popularity, geothermal energy still requires technological development and more efficient implementation [
3], primarily in areas with suitable geological conditions.
Geothermal energy is described as energy stored as heat beneath the ground inside the earth [
4]. According to several studies, the earth’s total heat emitted is around 42 × 10
12 W [
4]. Only 2% of this energy is present at the crust; the other 82% corresponds to the mantle due to the decomposition of several radioactive isotopes, and the remainder comes from the core. Despite the large amount of available energy, its utilization is limited to areas with specific geological conditions [
5].
In recent decades, technological advances have allowed for greater efficiency in geothermal heat extraction, reducing costs and expanding the geographic potential of this source [
6]. In addition, the focus on sustainability and reducing dependence on fossil fuels has led several governments to increase investment in geothermal research and development [
7].
Regarding electricity generation, geothermal energy’s total installed capacity in the twenty-first century’s first decades was nearly 9 MW. Nevertheless, the power associated with non-electrical uses was approximately 15 MW. In this sense, non-electrical applications involve space heating, greenhouses, aquaculture, industrial activities, or heat pumps, among others [
4].
Among geothermal solutions, ground source heat exchangers, which take advantage of the more stable subsurface temperature, are particularly promising for residential and commercial applications [
8]. This study focuses on a horizontal geothermal configuration implemented in a bioclimatic house.
Horizontal heat exchanger geothermal energy systems harness the thermal energy stored below the earth’s surface to provide efficient heating and cooling [
9]. These systems are mainly composed of pipes arranged horizontally at a relatively shallow depth, usually between 1 and 2 m [
10]. Although horizontal configurations are less expensive than vertical configurations [
11], their efficiency can be affected by seasonal thermal variability, which poses significant challenges in terms of energy efficiency [
12].
Different works focus on the importance of predicting thermal resistance or soil temperature in geothermal systems [
13,
14]. On the other hand, in ref. [
15], a fault detection and recovery approach is proposed to identify a malfunctioning sensor on a geothermal system. More specifically, artificial neural network (ANN)-based methods [
16,
17] or random forest techniques [
18] were applied to predict the performance of a geothermal heat pump system. Also, ref. [
19] develops a performance prediction model of an air-cooled heat pump system using ANN, support vector machine (SVM), random forest (RF), and K-nearest neighbor (KNN). In this sense, ref. [
20] proposes linear regression, nonlinear regression, and ANN techniques to predict the influence of the input variables on the heat transfer rate and heat transfer rate variation of a ground source heat pump (GSHP). In addition, ref. [
21] develops a method for predicting the behavior of a horizontal geothermal heat exchanger by means of several time series techniques.
As stated above, several approaches have been proposed for predicting the behavior of GSHPs or geothermal heat exchangers. Nevertheless, there is not extensive work focused on bioclimatic housing involving several renewable energy sources. In this sense, this research deals with a multi-energy context including wind power, solar thermal, solar photovoltaic, biomass, and geothermal energy systems. This is important to ensure an efficient system management and optimize the energy demand. Moreover, no published work in this context involves the use of the intelligent techniques proposed for this study.
The present research proposes a method based on intelligent regression techniques to predict the output temperature of a geothermal heat exchanger using the input temperature and the ground reference temperature as predictors. The exchanger includes these sensors and several buried temperature sensors throughout its length. One of the reasons for installing these buried sensors is to detect possible obstructions along the path. In this sense, this approach aims not only to optimize the thermal performance of the system and its energy demand but also to improve sensor maintenance, avoiding costly and time-consuming interventions.
In general terms, this study aims to promote more accessible housing by reducing the costs of sensor installation in difficult-to-access locations. It should be noted that the replacement of a buried probe can be expensive. In addition, the optimization of renewable energy systems by means of artificial intelligence (AI) techniques contributes to the reduction of greenhouse gas emissions. Furthermore, this work seeks to improve the performance of the methods used in previous work.
Satisfactory results were obtained by applying several machine learning (ML) techniques, achieving a good prediction of the output temperature of the heat exchanger. Specifically, recursive least square, K-nearest neighbors, decision tree, random forest, polynomial regression, support vector regression, and multilayer perceptron techniques were proposed for this study. In addition, several metrics were proposed for an in-depth model evaluation. Furthermore, a statistical analysis was carried out to determine which of the techniques provided the best results. Furthermore, this study has evidenced a significant correlation between different sensors of the geothermal system.
This paper is structured as follows. After the present introduction, a brief description of the case of study is presented. Then, the methods section is exposed. Subsequently, experiments and results are shown in the following section, and, finally, conclusions and future work are presented.
2. Case of Study
This section describes the geothermal system under study, located in a bioclimatic house, and the dataset obtained from periodical sensor measurements.
2.1. Sotavento Bioclimatic House
The bioclimatic house analyzed results from a project developed by the Sotavento Galicia Foundation, which aims to demonstrate the viability of different renewable systems and promote their use. This house is located in the Sotavento Experimental Wind Farm facilities between the municipalities of Xermade (Lugo) and Monfero (A Coruña) in the autonomous community of Galicia, Spain. Its geographical coordinates are 43°21′ North, 7°52′ West. It is located at an altitude of 640 m and 30 km from the sea.
The bioclimatic house’s thermal and electrical installations are supported by several types of renewable energy systems. Wind turbines and photovoltaic panels generate electricity. Furthermore, when these energy sources are insufficient to meet demand, the power grid supplies electricity to the home’s lighting and power systems. In contrast, the thermal installation manages the charge of the heating system and domestic hot water (DHW). Solar, geothermal, and biomass energies serve to carry out this task.
The thermal installation is divided into three main sections: generation, accumulation, and consumption.
Generation: The generation system is composed of three renewable energy sources:
Solar thermal system: The solar panels absorb energy from the solar radiation and heat the fluid of the primary circuit. Then, this fluid is fed into a solar accumulator.
Biomass boiler system: a biomass boiler with a pellet yield of 90% provides hot water to the inertial accumulator, ensuring an internal temperature of around 63 °C.
Geothermal system: It combines a ground source heat pump and a horizontal heat exchanger. The heat exchanger consists of several pipes arranged horizontally at depth of 2 m. The warm water from the heat pump is driven directly to the inertial accumulator.
Accumulation: this group comprises two accumulation units: a solar accumulator and an inertial accumulator for heat supply and DHW. A preheating system is also integrated to reach the temperature setpoint if necessary.
Consumption: The inertial accumulator supplies DHW and underfloor heating. The DHW supplies the bathroom and kitchen, dimensioning the system according to the Spanish Technical Building Code (240 L per day). The underfloor heating system keeps the house temperature between 18° C and 22 °C. For this purpose, the water temperature is regulated between 35 °C and 40 °C.
As stated in the Introduction section, this paper focuses on the geothermal energy system described in depth in the following subsection.
2.2. Geothermal Heat Pump and Horizontal Heat Exchanger
As stated above, heat generation through a horizontal heat exchanger is among the proposed active solutions.
Figure 1 shows the topology of the geothermal system, involving both heat pump (AT2) and horizontal heat exchanger. In this case, two circuits are connected to the heat pump: the primary circuit, including the heat exchanger (the circuit of the right side including S28 and S29 temperature sensors), and the second circuit (left side including S30 and S31 temperature sensors and a flowmeter C6), connecting the heat pump to the inertial accumulator.
Figure 2 shows a simplified scheme of the horizontal heat exchanger. The geothermal heat exchanger is composed of four parallel circuits with a total length of 100 m per circuit. Eight sensors (S3xx) are located along each circuit to measure the ground temperature in different positions while the system operates. Also, a reference sensor (S401) is used to monitor the ground temperature. Furthermore, the system includes input (S28 in
Figure 1) and output (S29 in
Figure 1) temperature sensors for the geothermal heat exchanger. Finally, a thermal power sensor is located at the second circuit of the heat pump.
2.3. Dataset Description
A sensor data acquisition over a 1 year period has been performed. It should be noted that parallel circuit sensor data are only available for one of the four circuits. In this sense, the dataset comprises a total of 52,645 observations with a sampling rate of 10 min. The features that compose the dataset are described below:
date-time (yyyy-mm-dd hh:mm): date-time for the corresponding observation.
Tin (°C): return temperature of the horizontal heat exchanger circuit to the heat pump.
Tout (°C): output temperature from the heat pump to the heat exchanger circuit.
Tref (°C): reference temperature of the ground.
c1sn (°C): temperature sensor n for horizontal heat exchanger circuit 1.
Pavg (W): output thermal power of the heat pump, averaged over 10 min.
2.4. Data Preparation
First, a data exploration stage was performed where several observations were discarded due to missing data. Then, from the initial 52,645 samples, the resulting number of samples is reduced to 52,638. Next, the samples where the heat pump was turned off, resulting in an output thermal power, or Pavg, of 0 W, were eliminated, yielding a dataset of 4883 samples. Then, the date-time feature was removed from the dataset. Finally, a standardization was applied to the dataset, giving a mean value of 0 and a variance of 1.
5. Conclusions and Future Work
In this paper, a machine learning-based approach has been proposed to predict the output temperature of a horizontal geothermal heat exchanger of a bioclimatic house. Considering that multiple sensors are installed under the ground, this methodology aims to reduce installation and maintenance costs by minimizing the number of sensors used to perform the prediction. In this sense, the models could also be used to predict obstructions inside the exchanger.
Other advantages of applying this methodology include detecting possible physical sensor reading deviations. Furthermore, it can be used to predict the behavior of the heat exchanger to give information to the overall energy management system of the bioclimatic house. Moreover, the intelligent techniques proposed in this study could be deployed in an edge computing context, thus optimizing energy resources.
The methodology proposed in this study can apply to other cases, employing a significant dataset that represents the behavior of a different system or heat exchanger. Hence, this approach could be used to assess the feasibility of future installations. Nevertheless, the present research is focused on a unique installation. Thus, the modeling was carried out for the exchanger installed in this specific bioclimatic house. In this sense, the model validation and test stages have been conducted using the available dataset.
Several ML regression techniques were applied, and promising results were achieved. Specifically, the MLP technique obtains a good prediction, committing only a mean absolute error of 0.4204 °C and a mean squared error of 0.3887 °C2 while obtaining a symmetric mean absolute percentage error of 4.6589%. The coefficient of determination also obtains a high value of 0.9308, indicating that the regression line fits well with the ground values.
Assessing the performance of a geothermal heat exchanger is essential for optimizing energy consumption. In this sense, several strategies can be developed to improve energy efficiency and foster the implementation of predictive maintenance strategies. In addition, future applications, such as proper temperature regulation or electrical smart grids, may align with this proposal. In addition, given the multi-energy design of the bioclimatic house, future work may focus on the overall management of the whole thermal energy system.
More specifically, other approaches could be evaluated for implementing virtual sensors, in line with cost and maintenance optimization, focusing on sensor malfunction detection. Furthermore, other statistical or deep learning techniques such as ARIMA, LSTM, or GRU could be proposed to perform temperature forecasting, extending the prediction horizon.