A predictive control method for a collaborative model of a district cooling system that spans the entire source-grid-load chain.

Abstract

This invention provides a predictive control method for a district cooling system based on a collaborative model spanning the entire power supply, grid, and load chain, belonging to the field of intelligent control technology for cooling systems. The method involves collecting and preprocessing meteorological and operational characteristic data of the system; inputting the data into an Attention-LSTM cooling load prediction model, determining the optimal step size using thermal inertia analysis, and outputting a cooling load prediction sequence; constructing a quasi-steady-state hydraulic-thermal coupling mechanism model of the pipe network based on unified energy path theory; establishing an optimization model with three objectives, including pump power consumption, using load prediction as input and the coupling model as constraints; employing a multi-objective particle swarm optimization algorithm to obtain the optimal operating parameters that satisfy the thermal imbalance constraint; and combining peak-valley electricity pricing to formulate a collaborative scheduling strategy for chiller units and cold storage devices, generating the final control command. This method effectively solves the problems of insufficient terminal cooling supply and localized overcooling caused by hydraulic imbalance and large hysteresis characteristics in the pipe network, achieving optimal control of operating costs.

Description

Technical Field

[0001] This invention belongs to the field of intelligent control technology for cooling systems, and particularly relates to a collaborative model predictive control method for a district cooling system that runs through the entire source-grid-load chain. Background Technology

[0002] District cooling systems typically consist of a cooling plant, a distribution network, and end users. The theoretical basis for their operation and control lies in achieving a thermal balance with end-user demand by adjusting the cooling plant's output, network flow rate, and temperature. Ideally, the system should operate with the lowest possible energy consumption and operating costs while still meeting users' thermal comfort requirements.

[0003] Currently, the control of district cooling systems largely relies on empirical value adjustment or simple PID feedback control. With the development of artificial intelligence, neural networks have begun to be applied to cooling load forecasting, and model predictive control (MPC) has also been gradually introduced to guide cooling source production by predicting load demand in advance.

[0004] Although existing technologies have made some progress in improving system stability, the following significant drawbacks still exist in practical engineering applications:

[0005] 1. Due to the large scale of district cooling system piping networks and significant transmission delays, existing regulation strategies often fail to accurately match the relationship between flow rate and temperature difference. This leads to the system frequently operating in an inefficient state of high flow rate and small temperature difference, significantly increasing the power consumption of distribution pumps and causing serious energy waste.

[0006] 2. Most existing forecasting methods rely purely on historical data, ignoring the inherent thermal inertia of building and piping systems. Due to the lack of physical constraint analysis on thermal inertia, the setting of the forecast time domain is often arbitrary, making it difficult to balance forecast accuracy and decision-making timeliness, and failing to provide truly accurate and timely feedforward signals for the control system;

[0007] 3. Existing technologies typically treat cooling stations, pipeline distribution, and terminal loads as independent components for optimization. Due to the lack of a coordinated mechanism for the entire system's "source-network-load" structure, and the computational bottleneck caused by the complexity of the hydraulic-thermal coupling model, existing solutions struggle to achieve globally optimal control in complex dynamic environments, thus limiting the further realization of the system's energy efficiency potential.

[0008] The main reason for the above problems is that the district cooling system is a typical nonlinear, large-time-lag, and strongly coupled complex system. The difficulty in solving these problems lies in two aspects: on the one hand, establishing a high-precision full-system hydraulic-thermal dynamic model will bring a huge computational burden, making it difficult to meet the requirements of real-time control; on the other hand, how to ensure the thermal comfort of end users while taking into account the economic benefits of electricity price fluctuations and the overall energy efficiency of the system under complex time-varying operating conditions is a multi-objective, highly coupled nonlinear optimization problem. Summary of the Invention

[0009] To address the above problems, this invention provides a predictive control method for a district cooling system based on a collaborative model that spans the entire source-grid-load chain, comprising the following steps:

[0010] S1: Collect meteorological and operational characteristic data of the district cooling system and preprocess them into model input data types;

[0011] S2 inputs the S1 data into the Attention-LSTM-based cooling load prediction model, and combines the system thermal inertia analysis to determine the optimal prediction step size, outputting the user cooling load prediction sequence in the future time domain;

[0012] S3, based on the unified energy path theory, abstracts the cooling pipe network into an equivalent circuit topology, establishes the mapping between pressure drop and flow rate in the hydraulic model, establishes the coupling equation between temperature difference and heat flow in the thermal model, and constructs a quasi-steady-state hydraulic-thermal coupling mechanism model of the pipe network.

[0013] S4. Using the load forecast sequence as input and the coupling mechanism model as constraint, the flow rate of the pipeline branch, the node temperature and the valve opening are selected as design variables to establish an optimization model that includes three objective functions: pump power consumption, total cooling load of the system and thermal imbalance on the user side.

[0014] S5. The multi-objective particle swarm optimization algorithm is used to perform rolling optimization on the established optimization model. The Pareto solution that satisfies the thermal imbalance threshold constraint and has the lowest comprehensive energy consumption is searched in the solution space to obtain the optimal pipeline network operation parameters.

[0015] S6, based on the optimal pipeline operation parameters and combined with real-time peak and valley electricity price information, formulates a coordinated scheduling strategy for the cold source side units and cold storage devices, prioritizing cold storage during valley electricity periods and prioritizing ice melting during peak periods, and generates the final control command.

[0016] Preferably, the meteorological features include outdoor dry-bulb temperature, wind speed, and precipitation; the operational features include pipeline flow rate, heat flux, and cooling load value; the collected raw data are cleaned to remove missing and outlier values, and the Min-Max normalization method is used to map all features to the [0,1] interval.

[0017] Preferably, the Attention-LSTM-based cold load prediction model includes an input layer, an LSTM extraction layer, an Attention weighting layer, and a fully connected output layer;

[0018] The LSTM extraction layer has a time step of 1. T Meteorological characteristics and operational characteristic sequences Input LSTM cells; LSTM cells contain forget gates. f t Input gate i t and output gate o t Used to update cell state C t And output the hidden layer state vector h t ；

[0019] The Attention weighted layer includes an LSTM layer. T The hidden state sequence output at each time step is: The attention mechanism utilizes a shared, learnable weight matrix. With each hidden state Multiply by the product to calculate an alignment score that measures the importance of different time steps to the final task. , This is the bias term. The score is then normalized using the Softmax function and converted into attention weights. Then, a context vector is generated by weighted summation of the hidden states. ;

[0020] Finally The input is a fully connected layer, which, after inverse normalization, outputs the predicted cooling load values ​​for future time periods. .

[0021] Preferably, the basic forecast step size is set to 20 minutes, and a multi-step rolling forecast strategy is adopted to output the load forecast sequence for the next 3 steps, i.e., the next 60 minutes, at once.

[0022] Preferably, the S3 process specifically includes:

[0023] Constructing a hydraulic equivalent network: targeting networks containing J Each pipe section and I The cooling pipe network at each node will reduce the pressure drop in the pipe section. Analogous to voltage, the square of volumetric flow rate Analogous to electric current, nonlinear flow resistance is defined based on the Darcy-Weisbach formula. Analogous to a resistor, the details are as follows:

[0024] ;

[0025] ;

[0026] in This refers to the equivalent absolute roughness of the pipe's inner wall. The density of the fluid inside the pipe, For pipe diameter, For the length of the pipe, For the equivalent length of local resistance, , Let be the local drag coefficient, when hour, ;

[0027] The flow distribution and pressure state of each pipe segment can be solved by simultaneously solving the flow balance and loop pressure drop balance equations of the equivalent network as follows, which can be used as the hydraulic boundary conditions for thermal calculations.

[0028] ;

[0029] in Let i be the traffic of the i-th node. I For a set of nodes, For the first Total loop voltage drop L For loop sets, Let i be the heat flux at node i. For the first Loop temperature drop and.

[0030] Constructing a thermally equivalent network: Addressing excess temperature at nodes Analogous to voltage, heat flux density Analogous to electric current; constructing thermal resistance that includes fluid convection heat transfer. Thermal resistance of the tube wall The series thermal resistance unit is as follows:

[0031] ;

[0032] ;

[0033] in, Specific heat capacity of the fluid For fluid mass flow rate, and These are the ambient temperature and the temperature of the fluid inside the pipe, respectively. The length of the differential pipe segment. This refers to the length of the pipe section.

[0034] The temperature of each node is determined by simultaneously solving the node flow balance and loop pressure drop balance equations of the equivalent network.

[0035] Preferably, in step S4, the flow rate of a pipe segment in the pipeline network is selected. Temperature of nodes And the user-side balancing valve Impedance as a design variable X Constructing a system that includes pump power Total cooling capacity of the system and user-side thermal imbalance The multi-objective optimization function is as follows:

[0036] ;

[0037] Pump power , For the motor safety factor, For pump flow rate, For the water pump head, For fluid density, and These are pump transfer efficiency and pump efficiency, respectively; user-side thermal imbalance. , For the temperature difference between the inlet and outlet water for users, The user's cooling demand; the total cooling load of the system. ;

[0038] Preferably, in step S4, a thermal imbalance threshold is set. As a core constraint, the flow rate of the pipeline section and the impedance of the balancing valve are also constrained to be within the physically permissible range.

[0039] Preferably, the S5 process specifically includes:

[0040] MOPSO Algorithm Initialization and Iteration: The established optimization model is solved using a multi-objective particle swarm optimization algorithm. The load prediction sequence output by S2 is used as input. The search in the solution space is performed by updating the particle velocity and position. The particle velocity update formula is as follows:

[0041] ;

[0042] The position update formula is as follows:

[0043] ;

[0044] in and The first k The velocity and position vectors of the particles; and These are the velocity and position vectors of the (k+1)th generation particle, respectively. Inertial weights; and For learning factors; For the individual optimal solution; This is the globally optimal solution. and It is the sum of random numbers in the interval [0,1].

[0045] Optimal operating point selection: After the algorithm converges, the Pareto front solution set is obtained. From this solution set, points satisfying the optimal operating conditions are selected. The solution is obtained, and the solution with the smallest normalized product of pump power, thermal imbalance and total cooling consumption is selected as the optimal operating point.

[0046] Preferably, the S6 process specifically includes:

[0047] Electricity price time period division: The system has a built-in local electricity price time period table, which is divided into off-peak electricity period, flat electricity period and peak electricity price period;

[0048] Coordinated scheduling strategy formulation: Based on the output of the optimal total cooling demand of the system, the following logic is executed: During off-peak hours, the refrigeration units are controlled to operate at full load, and the system prioritizes normal cooling supply according to the load. The remaining units are in cold storage mode to store ice energy, and the cold storage pump is turned on to store excess cold energy in the ice storage refrigerator; during normal hours, cooling is supplied normally according to the load; during peak electricity price periods, ice is melted, and the cold energy stored in the cold storage refrigerator is released first. At this time, the cold storage refrigerator and the unit share the load.

[0049] The determined unit start-up and shutdown status, ice storage and melting mode switching signals, water pump frequency and valve opening setpoints are packaged to generate the final control command vector.

[0050] Compared with the prior art, the present invention has the following beneficial effects:

[0051] The core of this invention lies in its overall logical concept, which incorporates the entire "source-network-load" chain of a district cooling system into a unified and collaborative framework. It utilizes an Attention-LSTM network to capture the nonlinear temporal characteristics of the load, and combines this with system thermal inertia analysis to determine the optimal prediction step size, thus solving the prediction lag problem under long transport delays. A unified energy path theory is introduced to transform the complex hydraulic-thermal coupling process of the pipe network into an equivalent circuit topology. Pipe flow rate, node temperature, and valve opening are converted into design variables, and pump consumption, cooling capacity, and thermal imbalance are used as target responses to establish a multi-objective optimization model, transforming actual operation and control into a mathematical optimization problem. A multi-objective particle swarm optimization algorithm is used for rolling solution, and based on the Pareto optimal solution set, a real-time peak-valley electricity pricing strategy is combined to make decisions, resulting in an optimal collaborative operation scheme that balances economy and comfort.

[0052] This invention provides a predictive control method for a district cooling system based on a collaborative model. This method effectively addresses the issues of insufficient cooling supply and localized overcooling at the terminal level caused by hydraulic imbalances and large time lags in traditional control strategies. While strictly ensuring thermal balance on the user side, it achieves optimal control of operating costs through electricity-price-guided cold storage and release scheduling, resulting in the best balance between thermal comfort, system energy efficiency, and operational economy. Compared to traditional CFD models, the method significantly improves online computational efficiency; compared to pure data models, it enhances physical interpretability and robustness. It is suitable for refined management of cooling systems under complex dynamic conditions, thereby improving the overall intelligent operation level of the system. Attached Figure Description

[0053] To more clearly illustrate the technical solutions of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the following description is only one embodiment of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0054] Figure 1 This is a flowchart illustrating the overall steps of the present invention.

[0055] Figure 2 This is a diagram of the overall control framework of the present invention.

[0056] Figure 3 This is a structural diagram of the Attention-LSTM-based cold load prediction model of this invention.

[0057] Figure 4 This is a structural diagram of the LSTM layer in the Attention-LSTM-based cold load prediction model of this invention.

[0058] Figure 5 This is a graph showing the evaluation metrics for each network model in the embodiments of the present invention.

[0059] Figure 6 This is a graph showing the model prediction error under different prediction step sizes in an embodiment of the present invention.

[0060] Figure 7 This is the pipeline network optimization result under typical working time in the embodiments of the present invention.

[0061] Figure 8 This is a comparison diagram of the system thermal imbalance under different strategies in the embodiments of the present invention.

[0062] Figure 9 This is a comparison chart of system economy and energy consumption under different strategies in the embodiments of the present invention.

[0063] Figure 10 This is a schematic diagram of the hardware structure in an embodiment of the present invention. Detailed Implementation

[0064] Please see Figure 1 and Figure 2 This invention proposes a collaborative model predictive control method for district cooling systems that spans the entire "source-grid-load" chain. This method constructs an Attention-LSTM load prediction model and a network coupling model based on energy path theory, and utilizes a multi-objective particle swarm optimization algorithm for collaborative optimization to achieve the optimal trade-off between system energy efficiency and terminal thermal comfort.

[0065] The system structure in this embodiment is as follows: Figure 10 As shown, a district cooling system mainly consists of three core components: the cold source side, as the core of the system, integrates the refrigeration unit, cooling tower, and energy storage device, responsible for centralized cooling capacity production; the distribution network forms a closed double-loop system, undertaking the transportation of chilled water from the station to users and the collection of return water; and the user side is the terminal of the system. It is typically separated from the main network by plate heat exchangers, forming a secondary circulation to meet the personalized air conditioning and cooling needs of buildings.

[0066] The present invention will be further described below with reference to embodiments. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0067] I. Establishing a cold load prediction model based on Attention-LSTM:

[0068] Historical meteorological data (temperature, humidity, radiation, etc.) and building operation data (supply and return water temperature, flow rate, etc.) of the district cooling system are collected. A Long Short-Term Memory (LSTM) network incorporating attention mechanisms is constructed, such as... Figure 3 , Figure 4 As shown.

[0069] The LSTM extraction layer has a time step of 1. T Meteorological characteristics and operational characteristic sequences Input LSTM cells; LSTM cells contain forget gates. f t Input gate i t and output gate o t Used to update cell state C t And output the hidden layer state vector h t；

[0070] The Attention weighted layer includes an LSTM layer. T The hidden state sequence output at each time step is: The attention mechanism utilizes a shared, learnable weight matrix. With each hidden state Multiply by the product to calculate an alignment score that measures the importance of different time steps to the final task. , This is the bias term; the score is then normalized using the Softmax function and converted into attention weights. Then, a context vector is generated by weighted summation of the hidden states. ;

[0071] Finally The input is a fully connected layer, which, after inverse normalization, outputs the predicted cooling load values ​​for future time periods. .

[0072] LSTM units are used to capture long-term dependencies in time-series data, and an attention mechanism is employed to calculate weight scores at different time steps, dynamically enhancing the influence of key moment features. The processed multidimensional feature sequence is then input into the model to establish a nonlinear mapping relationship between the input parameters and future cooling loads, thereby obtaining a load forecasting model with high dynamic capture capabilities.

[0073] The test data was selected from the case study test set, and the results are as follows: Figure 5 As shown, the model has a mean absolute percentage error of 5.746%, a root mean square error of 0.016, and a coefficient of determination R0. 2 The value reached 0.9996, indicating that the model has extremely high prediction accuracy and can accurately capture the dynamic changes in building load.

[0074] II. Determination of Optimal Prediction Step Size Based on System Thermal Inertia

[0075] To balance prediction error with the physical timeliness of control response, a comparative analysis of prediction step sizes with different time lengths was conducted. The mean absolute percentage error and root mean square error were calculated for different step sizes, and the nonlinear growth law of error with increasing step size was analyzed. Considering the heat capacity of the district cooling system network and the thermal inertia characteristics of the buildings, a time window (e.g., 20 minutes) with low error and covering the effective dynamic response period of the system was selected as the optimal prediction step size. A sequence of multiple consecutive step sizes was then used as the input sequence for subsequent optimization.

[0076] The calculation results in this embodiment are as follows: Figure 6As shown, the results indicate that the error increases non-linearly with increasing prediction step size. The model exhibits the lowest mean absolute percentage error (MASE) of only 1.86% and a root mean square error (RMS) of 267.50 kW when the prediction step size is 20 minutes (2 steps). Based on the analysis of the large network heat capacity and building thermal inertia of the district cooling system, 20 minutes was determined as the optimal prediction window for system stability. This step size ensures both high prediction accuracy and coverage of the system's effective dynamic response period; therefore, it was set as the optimal prediction step size, and three consecutive step sizes (60 minutes) were used as the input sequence for subsequent optimization.

[0077] III. Establishing a Quasi-steady-state hydraulic-thermal coupling model for the pipe network

[0078] Based on the unified energy path theory, the complex fluid transport and distribution network is abstracted into an equivalent circuit topology with potential flow resistance, as shown in Table 1.

[0079] Table 1. Correspondence Table of Unified Energy Path Theory

[0080]

[0081] In the hydraulic model, With power model By analogy, the pressure drop in a pipe section can be compared to voltage, the square of the flow rate to current, and the flow resistance, defined based on the Darcy-Weisbach formula, to resistance, as follows:

[0082] ;

[0083] in, This refers to the equivalent absolute roughness of the pipe's inner wall. For pipe diameter, For the length of the pipe, For the equivalent length of local resistance, , This is the local drag coefficient. Let be the coefficient of friction, when hour, .

[0084] Construct an equivalent hydraulic network and establish hydraulic loop equilibrium and nodal equilibrium equations:

[0085] ;

[0086] in Let i be the traffic of the i-th node. I For a set of nodes, For the first Total loop voltage drop L It is a set of loops.

[0087] The flow distribution within the pipeline network can be solved by solving the above equations.

[0088] In the thermodynamic model, a quasi-steady-state assumption is adopted, and excess temperature is first introduced. Calculate heat flux , , and Given the ambient temperature and the fluid temperature inside the pipe, respectively, the one-dimensional transport equation inside the pipe is used as the basic formula for the thermal circuit, as follows:

[0089] ;

[0090] in, Specific heat capacity of the fluid This refers to the fluid mass flow rate.

[0091] Since the control step size is 1 hour, and the transfer delay of cold energy in the cooling system is much less than 1 hour, the dynamic thermodynamics is simplified. Within each time step (1 hour), the thermodynamic process is regarded as a steady-state process. Therefore, the above formula can be simplified by eliminating the time variation term, resulting in the simplified dynamic analysis equation as follows:

[0092] ;

[0093] The simplified formula can be transformed to obtain the following formula:

[0094] ;

[0095] The simplified formula described above can be compared to Kirchhoff's laws for DC steady-state circuits, as follows:

[0096] ;

[0097] in, For current, For voltage, For electrical conductivity, It is a resistor.

[0098] The temperature difference can be likened to voltage, and the heat flux density to current, based on the above equations. A thermal resistance element is constructed based on convective heat transfer and pipe wall conduction. The thermal loop equilibrium and nodal equilibrium equations are then solved, as detailed below:

[0099] ;

[0100] A set of hydraulic-thermal coupled equations can be established to solve for the temperature distribution within the pipe network.

[0101] This model is used to calculate the transport delay and cycle delay of the pipeline network, ensuring that the time scale of the model calculation matches the control step size, thereby preserving key physical characteristics.

[0102] In this embodiment, a cooling network consisting of 12 pipe segments and 12 nodes is constructed, and a set of hydraulic-thermal coupled equations is established. Calculations show that the network transport delay in this case is approximately 26.4 minutes, and the cycle delay is approximately 53.9 minutes, which matches the control step size. This method effectively preserves key physical characteristics and simplifies the modeling and calculation process.

[0103] IV. Multi-objective collaborative optimization of operating conditions based on model prediction

[0104] The output load forecast sequence is used as a feedforward command, and the established quasi-steady-state model is used as a constraint. The flow rate of each pipe segment in the pipeline network, the temperature of each node, and the impedance of the balancing valve at the user end are selected as design variables. The total system energy consumption, pump power, and thermal imbalance at the end users are used as optimization objectives. A multi-objective optimization mathematical model that considers both system energy efficiency and end-user comfort is established, with the objective function as follows:

[0105] ;

[0106] Pump power , For the motor safety factor, and The flow rate and head of the pump. For fluid density, and These are pump drive efficiency and pump efficiency, respectively; thermal imbalance. , For users' cold needs.

[0107] Based on the above objective function and constraints, a multi-objective particle swarm optimization algorithm is used to perform rolling optimization on the established optimization model. Taking the load prediction sequence output by the Attention-LSTM-based cold load prediction model at its optimal step size as input, the search is performed in the solution space by updating the particle velocity and position. The particle velocity update formula is as follows:

[0108] ;

[0109] The position update formula is as follows:

[0110] ;

[0111] in and The first k The velocity and position vectors of the particles; and These are the velocity and position vectors of the (k+1)th generation particle, respectively. Inertial weights; and For learning factors; For the individual optimal solution; This is the globally optimal solution. and It is the sum of random numbers in the interval [0,1].

[0112] Optimal operating point selection: After the algorithm converges, the Pareto front solution set is obtained. From this solution set, points satisfying the optimal operating conditions are selected. The solution is obtained, and the solution with the smallest normalized product of pump power, thermal imbalance and total cooling consumption is selected as the optimal operating point.

[0113] In this embodiment, the flow rate of 12 pipe segments, the temperature of 12 nodes, and the impedance of the balancing valves at 4 user terminals are selected as design variables to establish a multi-objective optimization mathematical model that takes into account both system energy efficiency and end-user comfort. The model also incorporates the upper and lower limits of the pipe network flow rate and the physical constraints of equipment operating limits, as follows:

[0114] ;

[0115] in For the flow rate of pipe section 1, The temperature of node 1, To balance the impedance of the valve at the user end.

[0116] Taking the 12th hour of a typical day as an example, the optimization results are as follows: Figure 7 As shown, the optimized pump power is 75.05 kW, the thermal imbalance is 0.0883, and the total cooling load is 24881.84 kW. This operating point achieves a high degree of thermal equilibrium with low energy consumption. The optimization results for a typical workday are shown in Table 2.

[0117] Table 2 Optimization Results for Typical Workdays

[0118]

[0119] V. Implement a dual-timescale collaborative control strategy:

[0120] A control mode combining long and short time scales is adopted, centralizing system control authority on the long-scale cycle. At the start of each long-scale cycle, the start-up and shutdown of the chiller unit, the scheduling of cold storage, and the opening of pumps and valves are uniformly adjusted based on the optimal output parameters. The short-scale cycle only provides high-frequency load forecast updates and does not directly trigger equipment adjustments. By covering three short-scale cycles with a long-scale cycle, the control decisions can anticipate and balance load fluctuations within the long-scale cycle. This mode effectively limits the adjustment frequency of core equipment while accurately maintaining terminal comfort, ensuring system operational safety and hardware lifespan.

[0121] VI. Implementation of a peak-valley electricity price-oriented coordinated control strategy for cooling sources:

[0122] By combining real-time peak and off-peak electricity price information with the obtained optimal pipeline network setting parameters, a coordinated scheduling strategy for the cooling source side is formulated. During off-peak electricity price periods, the chiller units are controlled to operate at full load and activate the cold storage mode to store cooling capacity using off-peak electricity prices. During peak electricity price periods or peak hours, the cold storage devices are prioritized for ice melting and cooling, with the chiller units supplementing the remaining capacity. The optimized pipeline flow rate, temperature setpoint, and chiller start / stop signals are sent to the central control system to drive the coordinated operation of the chiller units, variable frequency pumps, and valves.

[0123] In this embodiment, during the off-peak electricity price period from 0:00 to 8:00, the chiller unit is controlled to operate at full load and activate the cold storage mode to store cold energy using off-peak electricity prices. During the peak electricity price periods from 11:00 to 12:00 and from 15:00 to 17:00, the cold storage device is prioritized for ice melting and cooling, with the chiller unit supplementing the shortfall. The optimized pipeline flow rate and temperature setpoint are used as the control targets for the chiller unit and its supporting equipment, driving the chiller unit, variable frequency pump, and valves to work in tandem, achieving a shift from passive response to proactive optimization.

[0124] Within each control step, the parameters of the prediction model and the pipeline network model are corrected using system measured feedback data. This eliminates accumulated errors caused by environmental disturbances or model simplification, and the prediction and optimization process is repeated at a new time starting point. Continuous rolling optimization ensures that the system is always in optimal operating condition, solving the problem of inefficient system operation.

[0125] VII. Simulation Verification and Performance Comparison:

[0126] The collaborative control strategy proposed in this invention was applied to a dynamic simulation model for verification and compared with the traditional demand-driven supply strategy. The case results are as follows: Figure 8 , Figure 9 As shown, Figure 8 This demonstrates the total thermal imbalance of users under different control strategies. Figure 9 It demonstrates the economic and energy consumption performance under different control strategies.

[0127] Appendix Figure 9 The three schemes corresponding to (a), (b), and (c) in the figure are Case 3 of the present invention, and the corresponding appendix is ​​attached. Figure 9 (c) Figure 9 (a) corresponds to Case 1. It only uses the load forecasting model to obtain user cooling loads, without undergoing multi-objective optimization at the pipeline network end. It adopts a direct demand-driven supply model. This scheme corresponds to the benchmark traditional scheme and is also supplemented by… Figure 8 The benchmark traditional scheme in the middle. Figure 9In section (b), corresponding to Case 2, the user's cooling load is obtained through the load prediction model, and multi-objective optimization is performed at the pipeline end. However, the collaborative model predictive control method described above is not used. Instead, the total cooling consumption after multi-objective optimization at the pipeline end is used as the benchmark, and cooling is implemented through a simple PID control mode. This scheme is an optimized and upgraded scheme, used for comparison with the strategy of this invention.

[0128] The results show that under the traditional strategy, the system thermal imbalance reaches a maximum of 6.188, indicating severe localized overcooling. After applying the strategy of this invention, the user-side thermal imbalance is strictly controlled within 0.16, and below 0.12 for most periods, significantly improving cooling quality. In terms of economics, although ensuring thermal balance leads to an 8.70% increase in total energy consumption, through electricity price-guided cold storage and release scheduling, the total daily operating cost only increases from 66,701.82 yuan to 68,762.69 yuan, with a cost increase controlled at 3.09%. This result verifies that this scheme achieves the optimal trade-off between thermal comfort and system energy efficiency with minimal economic cost.

[0129] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0130] While the specific embodiments of the present invention have been described above, they are not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A predictive control method for a district cooling system based on a collaborative model spanning the entire source-grid-load chain, characterized in that, The process includes the following: S1: Collect meteorological and operational characteristic data of the district cooling system and preprocess them into model input data types; S2 inputs the S1 data into the Attention-LSTM-based cooling load prediction model, and combines the system thermal inertia analysis to determine the optimal prediction step size, outputting the user cooling load prediction sequence in the future time domain; S3, based on the unified energy path theory, abstracts the cooling pipe network into an equivalent circuit topology, establishes the mapping between pressure drop and flow rate in the hydraulic model, establishes the coupling equation between temperature difference and heat flow in the thermal model, and constructs a quasi-steady-state hydraulic-thermal coupling mechanism model of the pipe network. S4. Using the load forecast sequence as input and the coupling mechanism model as constraint, the flow rate of the pipeline branch, the node temperature and the valve opening are selected as design variables to establish an optimization model that includes three objective functions: pump power consumption, total cooling load of the system and thermal imbalance on the user side. S5. The multi-objective particle swarm optimization algorithm is used to perform rolling optimization on the established optimization model. The Pareto solution that satisfies the thermal imbalance threshold constraint and has the lowest comprehensive energy consumption is searched in the solution space to obtain the optimal pipeline network operation parameters. S6, based on the optimal pipeline operation parameters and combined with real-time peak and valley electricity price information, formulates a coordinated scheduling strategy for the cold source side units and cold storage devices, prioritizing cold storage during valley electricity periods and prioritizing ice melting during peak periods, and generates the final control command.

2. The predictive control method for a district cooling system with a collaborative model spanning the entire source-grid-load chain as described in claim 1, characterized in that: The meteorological features include outdoor dry-bulb temperature, wind speed, and precipitation; the operational features include pipeline flow rate, heat flux, and cooling load; the collected raw data are cleaned to remove missing and outlier values, and the Min-Max normalization method is used to map all features to the [0,1] interval.

3. The predictive control method for a district cooling system with a collaborative model spanning the entire source-grid-load chain as described in claim 1, characterized in that: The cold load prediction model based on Attention-LSTM includes an input layer, an LSTM extraction layer, an Attention weighting layer, and a fully connected output layer. The LSTM extraction layer has a time step of 1. T Meteorological characteristics and operational characteristic sequences Input LSTM cells; LSTM cells contain forget gates. f t Input gate i t and output gate o t Used to update cell state C t And output the hidden layer state vector h t ； The Attention weighted layer includes an LSTM layer. T The hidden state sequence output at each time step is: ; Attention mechanisms utilize shared, learnable weight matrices With each hidden state Multiply by the product to calculate an alignment score that measures the importance of different time steps to the final task. , This is the bias term; the score is normalized using the Softmax function and then converted into attention weights. Then, a context vector is generated by weighted summation of the hidden states. ; Finally The input is a fully connected layer, which, after inverse normalization, outputs the predicted cooling load values ​​for future time periods. .

4. The predictive control method for a district cooling system with a collaborative model spanning the entire source-grid-load chain as described in claim 1, characterized in that: The basic forecast step size is set to 20 minutes, and a multi-step rolling forecast strategy is adopted to output the load forecast sequence for the next 3 steps, i.e., the next 60 minutes, at once.

5. The predictive control method for a district cooling system with a collaborative model spanning the entire source-grid-load chain as described in claim 1, characterized in that: The specific process of S3 includes: Constructing a hydraulic equivalent network: targeting networks containing J Each pipe section and I The cooling pipe network at each node will reduce the pressure drop in the pipe section. Analogous to voltage, the square of volumetric flow rate Analogous to electric current, nonlinear flow resistance is defined based on the Darcy-Weisbach formula. Analogous to a resistor, the details are as follows: ； ； in This refers to the equivalent absolute roughness of the pipe's inner wall. The density of the fluid inside the pipe, For pipe diameter, For the length of the pipe, For the equivalent length of local resistance, , Let be the local drag coefficient, when hour, ; The flow distribution and pressure state of each pipe segment can be solved by simultaneously solving the flow balance and loop pressure drop balance equations of the equivalent network as follows, which can be used as the hydraulic boundary conditions for thermal calculations. ； in Let i be the traffic of the i-th node. I For a set of nodes, For the first Total loop voltage drop L For loop sets, Let i be the heat flux at node i. For the first Loop temperature drop and; Constructing a thermally equivalent network: Addressing excess temperature at nodes Analogous to voltage, heat flux density Analogous to electric current; constructing thermal resistance that includes fluid convection heat transfer. Thermal resistance of the tube wall The series thermal resistance unit is as follows: ； ； in, Specific heat capacity of the fluid For fluid mass flow rate, and These are the ambient temperature and the temperature of the fluid inside the pipe, respectively. The length of the differential pipe segment. This refers to the length of the pipe section. The temperature of each node is determined by simultaneously solving the node flow balance and loop pressure drop balance equations of the equivalent network.

6. The predictive control method for a district cooling system with a collaborative model spanning the entire source-grid-load chain as described in claim 1, characterized in that: In step S4, the flow rate of a pipe segment in the pipeline network is selected. Temperature of nodes And the user-side balancing valve Impedance as a design variable X Constructing a system that includes pump power Total cooling capacity of the system and user-side thermal imbalance The multi-objective optimization function is as follows: ； Pump power , For the motor safety factor, For pump flow rate, For the water pump head, For fluid density, and These are pump transfer efficiency and pump efficiency, respectively; user-side thermal imbalance. , For the temperature difference between the inlet and outlet water for users, The user's cooling demand; the total cooling load of the system. , For the flow rate of pipe section 1, The temperature of node 1, The temperature of node 6.

7. The predictive control method for a district cooling system with a collaborative model spanning the entire source-grid-load chain as described in claim 6, characterized in that: In step S4, a thermal imbalance threshold is set. As a core constraint, the flow rate of the pipeline section and the impedance of the balancing valve are also constrained to be within the physically permissible range.

8. The predictive control method for a district cooling system with a collaborative model spanning the entire source-grid-load chain as described in claim 1, characterized in that: The specific process of S5 includes: MOPSO Algorithm Initialization and Iteration: The multi-objective particle swarm optimization algorithm is used to solve the established optimization model. The load prediction sequence output by S2 is used as input. The search in the solution space is performed by updating the particle velocity and position. The particle velocity update formula is as follows: ； The position update formula is as follows: ； in and The first k The velocity and position vectors of the particles; and These are the velocity and position vectors of the (k+1)th generation particle, respectively. Inertial weight; and For learning factors; For the individual optimal solution; This is the globally optimal solution. and It is the sum of random numbers in the interval [0,1]. Optimal operating point selection: After the algorithm converges, the Pareto front solution set is obtained. From this solution set, points satisfying the optimal operating conditions are selected. The solution is obtained, and the solution with the smallest normalized product of pump power, thermal imbalance and total cooling consumption is selected as the optimal operating point.

9. The predictive control method for a district cooling system with a collaborative model spanning the entire source-grid-load chain as described in claim 1, characterized in that: The specific process of S6 includes: Electricity price time period division: The system has a built-in local electricity price time period table, which is divided into off-peak electricity period, flat electricity period and peak electricity price period; Coordinated scheduling strategy formulation: Based on the output of the optimal total cooling demand of the system, the following logic is executed: During off-peak hours, the refrigeration units are controlled to operate at full load, and the system prioritizes normal cooling supply according to the load. The remaining units are in cold storage mode to store ice energy, and the cold storage pump is turned on to store excess cold energy in the ice storage refrigerator; during normal hours, cooling is supplied normally according to the load; during peak electricity price periods, ice is melted, and the cold energy stored in the cold storage refrigerator is released first. At this time, the cold storage refrigerator and the unit share the load. The determined unit start-up and shutdown status, ice storage and melting mode switching signals, water pump frequency and valve opening setpoints are packaged to generate the final control command vector.

Images