1. Introduction
Based on statistics from the International Energy Agency (IEA), the building sector is among the top three energy-consuming industries, representing 36% of global energy consumption. In this sector, heating, ventilation, and air conditioning (HVAC) systems account for 65% of the energy utilized in buildings [
1]. Additionally, buildings generated approximately 39% of global greenhouse gas emissions, highlighting the importance of optimizing HVAC systems for energy efficiency. In practice, HVAC systems typically exhibited large peak power loads and low energy efficiency, with peak-load periods coinciding with urban power peak-load times. This led to grid undersupply during peak loads and oversupply during valley loads. The IEA identified district cooling as a critical energy sector requiring significant action to achieve net-zero emissions by 2050 [
2].
District cooling has become an appealing alternative to traditional cooling systems in areas with high energy density such as urban environments [
3]. The implementation and advancement of district cooling systems (DCSs) rely on the use of local renewable energy sources and natural resources [
4]. These systems primarily rely on a centralized cooling station to generate cold energy, which is then transmitted to end users through pipelines. Due to the integration of local renewable energy sources, DCS has proved to be more efficient than standalone heating and cooling systems [
5]. Most current research on district cooling systems focuses on technical and economic optimization. Chan et al. [
6] employed a genetic algorithm to determine the optimal configuration of a DCS pipe network, while Dorotich et al. [
7] developed an hour-based optimization model for district cooling and heating systems with CO
2 emissions and operating costs as the objective functions. Their results indicated that, for the same CO
2 emissions, the operating cost of district cooling systems was lower than that of standalone cooling systems. Furthermore, extensive literature has addressed the control of district heating and cooling systems [
8]. Jingzhao et al. [
9] proposed a predictive-control model for district cooling and heating systems based on load forecasting and time delay, and actual measurements at Tianjin University demonstrated that this model not only improved supply–demand matching but also reduced total energy consumption. Laura et al. [
10] developed a model predictive control approach for a district cooling system with compressors and absorption chillers to maximize efficiency. Simulation-based experiments showed that this energy-saving method achieved up to 50% savings. Wen Jie et al. [
11] presented a robust optimization method to reduce cooling-water system indeterminacy during the design and operation phases of a district cooling system [
8].
To reduce the impact of HVAC system operation on the power grid and ensure the stable operation of the power system [
8], it is essential to integrate large-scale regional cooling and heating systems with renewable energy and ice-storage technologies [
12,
13]. This integration helps reduce the peak load on the power system. Additionally, establishing an integrated predictive control model for district cooling systems that combines renewable energy and ice-storage technology is crucial for ensuring efficient and economic operation [
14].
Energy demand forecasting is crucial for promoting resource optimization and controlling DCS; therefore, accurate and rapid district cooling system load forecasting is necessary [
15]. Improving the performance of the prediction model was a key step in developing energy-saving control strategies for air-conditioning systems, ensuring their economical and efficient operation. In recent years, researchers worldwide have made significant advancements in predicting building heating and cooling loads, utilizing a variety of load forecasting methods such as physical modeling, empirical methods, regression models, and artificial intelligence techniques.
The traditional method for simulating and estimating building cooling loads was based on the building energy modeling programs (BEMPs). The research and development of thermal-process-based BEMPs dates back to the 1980s and resulted in the creation of widely used software such as DeST [
16], EnergyPlus [
17], and ESP-r [
18]. However, using BEMP for cooling load estimation required detailed input on building thermal performance and operating schedules, often necessitating calibration [
19,
20]. Due to the substantial workload involved, this method was often unsuitable for existing buildings, and the lack of precise inputs reduced prediction accuracy.
Most current research employs data-driven machine-learning algorithms for cooling load prediction. In reference [
21], a short-period multi-step prediction model, based on support vector machines and discrete wavelet transforms, along with a new method for district heating system (DHS) user heat load prediction, was proposed. In reference [
22], a short-period heat load prediction algorithm based on a feature fusion long short-term memory model (FFLSTM) was established for DHS optimization and control. Kwok et al. [
23] used an artificial neural network (ANN) model to predict energy use in office buildings in Hong Kong, achieving a best RMSE of 11.41%. These experiments demonstrated that while algorithms have advantages in forecasting accuracy, many load forecasting models still face shortcomings in terms of computing time and accuracy requirements.
Furthermore, previous studies have shown that BP neural networks possess strong nonlinear mapping, self-learning, generalization, and fault tolerance capabilities, enabling them to achieve better load prediction results [
24]. However, due to the influence of the gradient descent algorithm, weights and thresholds are generated randomly when using BP neural networks [
25]. This randomness can cause the algorithm to easily fall into local optima during operation, reducing the accuracy of the output results and affecting the established prediction model [
25]. Huang et al. [
26] proposed a universal neural embedding initialization framework to address the initialization problem in neural network models. However, while advanced initialization methods may offer performance improvements, they often require additional computational costs or complex parameter tuning, which increases the time and computational resources needed for model training. Furthermore, the practical effectiveness of these methods may be influenced by specific application scenarios, datasets, and network architectures. Therefore, the demand prediction of large-scale regional cooling systems using the BP algorithm combined with ice-storage technology requires further study.
Additionally, ensuring the smooth operation of the power grid [
27], minimizing the impact of end users on the grid, and integrating renewable energy sources often involves systems combining heat pumps and heat storage [
28]. In the optimization and control of energy storage systems, major methods such as model predictive control (MPC), supervised control, and machine learning control are commonly adopted [
8,
29]. For optimizing ice-storage systems, the literature commonly reports factors like constant priority, fixed scheduling, refrigerator priority, cold-storage priority [
30,
31,
32], and rule-based optimal control strategies. These strategies have been widely applied to optimize the coefficient of performance (COP) of heat pumps and heat-storage systems, yielding satisfactory results [
33,
34]. Sun et al. [
35] reviewed various heat-storage control strategies and concluded that optimization strategies are more cost-effective than conventional methods. Luo et al. [
36] conducted a modeling study on integrating air-conditioning and ITS systems in a shopping center, demonstrating an 11.3% daily operating cost reduction through effective optimization and regulation. Wei Qinglai et al. [
37] employed the data-driven adaptive dynamic programming (ADP) method to optimize the control of a cold-storage air conditioner, achieving significant operational cost reductions. Powell et al. [
38] utilized mathematical models to optimize coolers and cold stores, treating coolers as a single optimal unit for dynamic optimization approaches. They integrated neural network load forecasting technology, effectively reducing control costs. Chen et al. [
39] proposed a dynamic programming-based control scheme to determine the optimal capacity ratio of refrigeration and ice-storage equipment. However, there remains an urgent need to fully utilize renewable energy sources such as rivers and lakes and to develop a comprehensive predictive control model for regional refrigeration systems that integrate ice-storage technology and renewable energy.
In summary, accurately predicting the load demand of large regional cooling systems and optimizing their control through the integration of renewable energy and ice-storage technology is of paramount importance. Recent studies [
40,
41] indicated that combining the ABC algorithm with the BP neural network model significantly enhanced performance and convergence speed. The global search capability of the ABC algorithm helped prevent the BP network from getting stuck in local optima, enabling it to find more optimal weights and biases. Additionally, it optimized the initial parameter settings, accelerated network convergence, and improved training stability. Furthermore, the ABC algorithm dynamically adjusted the learning rate based on network conditions to control the training speed and stability, mitigating issues like oscillation or overfitting. By employing diversified weight updating strategies, the ABC algorithm enhanced network diversity, thereby improving model generalization and adaptability.
In this paper, an integrated predictive control model of a regional cooling system combining river source heat pumps and ice-storage devices was proposed. The structure and parameters of the ABC-BP load forecasting algorithm were determined through the comprehensive analysis of test data. An optimization scheme for the combined heat pump and ice storage regional cooling system was developed, focusing on minimizing user-side operation costs and system power consumption during peak electricity prices. The model’s accuracy was validated using experimental data, demonstrating its effectiveness in peak-shifting and valley-filling strategies for the power grid, thus alleviating grid pressure.
4. Results and Discussion
4.1. Forecasting Results and Performance Analysis of ABC-BP and BP Networks
For visualization, the forecasted and observed values were compared during the summer months from 19–22 August, as shown in
Figure 6. It can be observed that the forecasted value of the ABC-BP model was closer to the measured data, with a smaller error compared with the BP model. This improvement was attributed to the enhanced optimization mechanism of the ABC algorithm, which, inspired by the foraging behavior of bees, effectively explored the solution space to identify better initial weights and avoid local minima. Consequently, the ABC-BP model achieved more accurate predictions, demonstrating its ability to better describe the nonlinear relationship between the eight input characteristic values and the load demand. This result highlights the effectiveness of the ABC-BP approach and its potential for widespread use in building performance prediction.
Furthermore, the discrepancy in prediction errors around the value of 75, where the ABC-BP model showed an error of approximately +50% and the BP model showed an error of about −50%, likely occurred due to the differing sensitivities of the models to specific data ranges. The ABC-BP model, which optimized initial weights through the artificial bee colony algorithm, generally performed better but struggled with certain data ranges, leading to overestimation. Conversely, the BP model, reliant on gradient descent, had less consistent performance across different data ranges, resulting in underestimation in this region. This highlights the need for further refinement in both models to handle specific edge cases more effectively.
4.2. Load Distribution Results under Each Control Strategy on a Typical Day
Under the typical daily and hourly loads predicted by the ABC-BP algorithm, the proposed optimization strategy, chiller priority, ice-storage priority, and constant-proportion control strategy were simulated, and the cooling capacity distribution results are shown in
Figure 7.
Under the optimization strategy, during the electricity-price flat period (8:00–11:00, 17:00–20:00, and 22:00–24:00), the ice-melting cooling ratio was significantly small, approximately 0.4. Even from 18:00 to 20:00, the RSHP was used for cooling to reserve sufficient ice volume for the peak time of electricity price at night. During the entire cooling process of the clod-storage device, the cooling supply was concentrated during the peak segment of electricity price (11:00–17:00 and 20:00–22:00). During the peak daytime electricity price (11:00–17:00) period, the ice-melting cooling ratio was kept above 0.85, and during this period, only one RSHP was required to be turned on to satisfy the cooling demand. In the peak period at night (20:00–22:00), the use of single melting ice for cooling could effectively avoid the opening of the refrigeration unit during the peak hours.
4.3. Energy Consumption Analysis of Each Strategy on a Typical Day
Figure 8 shows the hourly operating power under the four operating strategies on a typical day. As observed in
Figure 8, during the peak segment of the electricity price, the consumption of power with the optimized control strategy is always maintained at the lowest level, and its power consumption is 2309.8 kW. Compared with the chiller priority, ice-storage priority, and constant proportion, the power consumption of the optimization strategy was reduced by 75.59%, 27.96%, and 22.94%, respectively. It can be observed that the optimization strategy can reduce the superposition of the grid peak load better. In the low-electricity-price stage, the optimization strategy, ice-storage priority, and constant proportion need to be iced. However, a sudden change occurred at 7 o’clock because the terminal load was large at 7 o’clock. Furthermore, it was in the ice-storage stage at this time, and the melting ice cannot be used for cooling; therefore, the RSHP was required to be turned on for cooling. Therefore, at this time, the immediate power consumption of the optimization strategy, constant proportion, and ice-storage priority control strategy reached their maximum values.
Table 6 summarizes the daily total power consumption and power consumption at each stage of each operation strategy. From the perspective of gross power consumption, under this load condition, the chiller priority control strategy was equivalent to the traditional control strategy without an ice-storage device. This eliminated the energy loss of ice storage; thus, the gross consumption of energy was the smallest. Because the optimization strategy optimized the chilled-water outlet temperature of the RSHP and kept the unit running at a better partial-load rate, it reduced the consumption of energy by 0.14% and 6.89%, respectively, when compared with the ice-storage priority and the constant proportional control strategy.
In addition, to better represent the ability of each operating strategy to shift the power consumption of the peak section, we introduced the peak shift rate
, which is expressed as follows:
where
and
represent the power consumption of the peak power segment with and without the cold-storage system, respectively.
Considering the chiller priority control strategy as the benchmark, the peak-shift rates of the optimization strategy, ice-storage priority, and constant proportion were 4.29, 1.18, and 1.07, respectively. It can be seen that the optimization strategy has the strongest peak-shaving ability and can significantly reduce power consumption during the peak period such that the maximum power required is reduced.
4.4. Operating Costs of Various Strategies on a Typical Day
Table 7 compares and analyzes the daily operating costs under the different control strategies. Evidently, compared with the three control schemes of the chiller priority, ice-storage priority, and constant proportion, the cost reduction of the optimization strategy was 40.30%, 10.54%, and 19.00%, respectively. During the peak segment of the electricity price, the total melting-ice volume of the optimized control strategy accounted for 79.0% of the total daily ice-melting and cooling capacity; therefore, it could transfer the peak load to the electricity price valley to the greatest extent.
4.5. Operating Analysis of Optimization Strategy under Other Typical Design Conditions
Table 8 summarizes the operating costs, energy consumption, and maximum power during the peak periods for different strategies at 25–100% of the design load. The use of both kWh and kW in
Table 8 serves specific purposes: kWh represents the total energy consumption over time, which is crucial for evaluating overall efficiency and cost-effectiveness; kW indicates the maximum power demand during peak periods, which is essential for managing peak loads and optimizing operating costs. Including both measures provides a comprehensive understanding of the system’s performance under different strategies.
Except for the host priority strategy, in terms of the average total energy consumption under the design load, compared with the ice-melting priority control strategy and the constant-proportion control strategy, the average energy-saving percentages of the optimization strategy were 1.025% and 1.775%, respectively. Simultaneously, in terms of the average cost savings under the design load, compared with the chiller priority, ice-storage priority, and constant-proportion control strategies, the average cost-saving percentages of the optimization strategy were 19.2%, 4.4%, and 5.1%, respectively. In terms of the maximum daily power consumption, compared with the chiller priority, ice-storage priority, and constant-proportion control strategies, the average power-saving percentages of the optimization strategy were 30.0%, 18.1%, and 8.9%, respectively. In addition, as the load rate decreased, the optimized operation strategy could achieve a better goal of reducing the maximum power demand in the peak power price segment. It can be observed that under various load conditions, the operating electricity cost of the optimization strategy and the maximum power in the peak electricity price segment were the lowest. Therefore, the optimization strategy could achieve better control effects.