A park multi-energy collaborative comprehensive energy supply control method based on carbon constraint

Abstract

The application discloses a kind of park multi-energy collaborative comprehensive energy supply control methods based on carbon constraint, comprising: according to power grid carbon intensity, time-of-use electricity price and renewable energy output prediction, construct three-dimensional carbon potential energy surface, identify carbon potential peak and valley, determine low carbon potential energy valley period and high carbon potential energy peak period;In high carbon potential energy peak period, through mixed integer linear programming model, with the weighted sum of operating cost and carbon emission cost minimum as objective function, generate the carbon economy benchmark trajectory of future first cycle;With benchmark trajectory as target, through model predictive control to carry out second cycle rolling optimization, output correction control instruction;Real-time monitoring carbon intensity mutation, trigger the output adjustment of depth reinforcement learning agent to be superimposed to correction instruction, simultaneously according to carbon potential energy elasticity coefficient trigger advance collaborative mode.In the guarantee system operation economy, significantly reduce the power grid carbon emission of high carbon period, improve renewable energy consumption level.

Description

Technical Field

[0001] This invention belongs to the field of multi-energy collaborative control technology in industrial parks, specifically relating to a carbon-constrained multi-energy collaborative integrated energy supply control method for industrial parks. Background Technology

[0002] As a multi-energy coupling application scenario directly facing end users, the integrated energy system of the industrial park is considered one of the important paths to achieve a low-carbon energy transition due to its advantages such as multi-energy complementarity and energy cascade utilization. The industrial park energy supply system typically integrates multiple energy devices such as photovoltaics, energy storage, and ground source heat pumps, which not only meet the diverse energy needs of users, but also improve the renewable energy absorption rate and the economic efficiency of system operation.

[0003] Some studies incorporate carbon emission costs into the total operating cost of the system by establishing carbon trading models, thereby constraining system carbon emissions through economic means. Another type of research focuses on analyzing the coupling mechanism between carbon and energy flows, constructing carbon-energy synergy hub models to analyze the synergistic effects of carbon and energy flows from a planning perspective, aiming to improve investment effectiveness and carbon reduction flexibility.

[0004] However, most studies incorporate carbon emissions as a cost component into optimization objectives, using carbon trading mechanisms or carbon quota constraints for scheduling decisions. When the power grid is in a high-carbon period, i.e., when the proportion of dirty electricity is high, it is difficult to substantially reduce carbon emissions in Scope II. Especially when dealing with sudden events such as abrupt changes in carbon intensity, the response speed and coordination of existing methods need improvement, making it difficult to achieve rapid carbon reduction while ensuring long-term economic viability.

[0005] Therefore, how to achieve precise control of carbon-energy synergy and effectively reduce carbon emissions in advance has become a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] This invention provides a carbon-constrained multi-energy coordinated integrated energy supply control method for industrial parks, which solves the problems of difficulty in balancing global economic efficiency and real-time rapid response, and difficulty in substantially reducing carbon emissions due to imperfect control architecture. While ensuring the economic efficiency of system operation, it significantly reduces the carbon emissions of grid electricity purchases during high-carbon periods, improves the renewable energy consumption rate, and achieves coordinated and optimized scheduling.

[0007] The technical solution adopted in this invention is as follows: A carbon-constrained multi-energy coordinated integrated energy supply control method for industrial parks, comprising: Based on the prediction of grid carbon intensity, time-of-use electricity price and renewable energy output, a three-dimensional carbon potential energy surface is constructed to identify carbon potential peaks and valleys and determine the low carbon potential energy valley period and the high carbon potential energy peak period. During the peak period of high carbon potential energy, a mixed integer linear programming model is used to generate the carbon economic baseline trajectory for the first future cycle with the objective function of minimizing the weighted sum of operating costs and carbon emission costs. Using the baseline trajectory as the target, a second-cycle rolling optimization is performed through model predictive control, and a correction control command is output. The duration of the second cycle is shorter than that of the first cycle. Real-time monitoring of sudden changes in carbon intensity triggers the deep reinforcement learning agent to output adjustment amounts superimposed on the correction instructions. Simultaneously, based on the carbon potential energy elasticity coefficient, an advanced collaborative mode is triggered to control the ground source heat pump to store heat in advance and discharge energy to replace grid power supply.

[0008] The carbon-constrained multi-energy coordinated integrated energy supply control method for industrial parks adopted in this invention also has the following additional technical features: Constructing a three-dimensional carbon potential energy surface and identifying carbon potential peaks and valleys includes: A three-dimensional mesh surface was constructed with time as the horizontal axis, power as the vertical axis, and carbon intensity as the vertical axis. The carbon potential energy function value was calculated at each mesh point. The gradient descent algorithm is used to search for local minima and local maxima on the surface to determine the peaks and valleys of the carbon potential. The low carbon potential valley period and the high carbon potential peak period are determined based on the carbon potential peak and valley.

[0009] Based on the aforementioned carbon potential peaks and valleys, the periods of low-carbon potential valleys and high-carbon potential peaks are determined as follows: A first threshold is determined based on multiple local minima of the surface, and a second threshold is determined based on multiple local maxima of the surface. Identify low carbon potential energy valley periods when carbon intensity is below the first threshold and high carbon potential energy peak periods when carbon intensity is above the second threshold.

[0010] The carbon economy benchmark trajectory includes at least the energy storage charging and discharging benchmark power, the ground source heat pump benchmark output, and the grid power purchase benchmark power.

[0011] The constraints of the mixed-integer linear programming model include power balance, heat balance, energy storage system state constraints, ground source heat pump operation constraints, and power grid interaction constraints. The carbon economy baseline trajectory is obtained based on grid carbon intensity, time-of-use electricity prices, and renewable energy output forecasts, combined with system equipment parameters.

[0012] The second-cycle rolling optimization is performed using model predictive control, specifically as follows: A model predictive controller is constructed using the aforementioned carbon economy baseline trajectory as the tracking target; Real-time tracking deviation and control rate of change are obtained, and the unfavorable baseline value is obtained. When the baseline unfavorable value is greater than the third threshold, the system is reprogrammed using a mixed-integer linear programming model to obtain a corrected control command.

[0013] Real-time monitoring of sudden changes in carbon intensity triggers the deep reinforcement learning agent to output an adjustment amount that is superimposed on the correction instruction, specifically as follows: Real-time monitoring of sudden changes in the carbon intensity of the power grid; when the rate of change in carbon intensity exceeds a set threshold, a deep reinforcement learning agent is triggered.

[0014] Triggering a deep reinforcement learning agent, specifically: Using the current tracking deviation and correction control command as input, the output adjustment amount is superimposed on the correction control command to generate the final execution command.

[0015] The advanced collaborative mode is triggered based on the carbon potential energy elasticity coefficient, specifically as follows: The carbon potential energy elasticity coefficient is calculated in real time. This coefficient is jointly determined by the energy storage state of charge, the current energy efficiency ratio of the ground source heat pump, and the real-time carbon intensity of the power grid. When the real-time carbon intensity of the power grid exceeds the fourth threshold and the carbon potential elasticity coefficient exceeds the fifth threshold, the advanced coordination mode is triggered.

[0016] A second aspect of the present invention provides a computing device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the method.

[0017] Due to the adoption of the above technical solution, the beneficial effects achieved by this invention are as follows: 1. In this invention, by constructing a three-dimensional carbon potential energy surface and identifying carbon potential peaks and valleys, the abstract power grid carbon intensity factor is transformed into a carbon potential difference-driven energy flow self-organizing optimization, thus solving the problem of the disconnect between carbon constraints and energy dispatch. Based on carbon potential energy field theory, it can automatically identify periods with a high proportion of dirty electricity in the power grid, providing clear directional guidance for subsequent energy dispatch and enabling carbon constraints to drive energy optimization.

[0018] Furthermore, by constructing a first-cycle baseline trajectory generation, a second-cycle rolling optimization, and a real-time carbon intensity mutation monitoring and response system, a multi-timescale hierarchical control approach is adopted, solving the technical challenge of balancing global economic efficiency with real-time rapid response. The first-cycle layer uses a mixed-integer linear programming model, with the objective function of minimizing the weighted sum of operating costs and carbon emission costs, to generate the carbon economic baseline trajectory for the future first cycle, ensuring global economic efficiency. The second-cycle layer uses this baseline trajectory as the objective, employing model predictive control for second-cycle rolling optimization, outputting corrective control commands to effectively address prediction deviations such as fluctuations in photovoltaic power output. The real-time layer monitors carbon intensity mutations, triggering a deep reinforcement learning agent to output adjustments that are superimposed on the corrective commands, resulting in a short response time. This three-layer architecture achieves an organic unity between long-cycle economic planning and short-cycle rapid response, avoiding the drawbacks of single-timescale control that may compromise overall performance.

[0019] Moreover, by introducing a carbon potential elasticity coefficient and triggering an advanced collaborative mode, the industry pain point of being unable to substantially reduce carbon emissions during high-carbon periods is addressed. Real-time calculation of the carbon potential elasticity coefficient quantifies the potential for coordinated scheduling of energy storage and heat pumps. This combined mode of heat pump preheating and energy storage, along with energy storage discharge substitution, shifts electricity demand during high-carbon periods to low-carbon periods in advance through thermal energy storage, achieving coordinated carbon reduction across time periods and energy types.

[0020] In summary, this invention constructs a complete carbon energy collaborative optimization control method, which significantly reduces the carbon emissions from grid power purchases during high-carbon periods while ensuring operational economy, improves the level of renewable energy consumption, and facilitates the low-carbon operation of the park's integrated energy system. Attached Figure Description

[0021] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a flowchart illustrating the carbon-constrained multi-energy coordinated integrated energy supply control method for industrial parks according to one embodiment of the present invention. Detailed Implementation

[0022] To more clearly illustrate the overall concept of the present invention, a detailed description will be provided below with reference to the accompanying drawings and examples.

[0023] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0024] like Figure 1As shown, a carbon-constrained multi-energy coordinated integrated energy supply control method for industrial parks includes: S100: Based on the prediction of grid carbon intensity, time-of-use electricity price and renewable energy output, a three-dimensional carbon potential energy surface is constructed to identify carbon potential peaks and valleys and determine the low carbon potential energy valley period and the high carbon potential energy peak period.

[0025] The main purpose of this step is to transform the abstract, time-varying grid carbon intensity factor into a carbon potential energy field, thereby providing an intuitive basis for subsequent energy dispatch decisions. Traditional methods typically incorporate carbon emissions as a cost item into the optimization objective, making dispatch decisions through carbon trading mechanisms or carbon quota constraints, but the physical meaning of carbon-energy coupling remains unclear.

[0026] First, obtain the grid carbon intensity forecast series, time-of-use electricity price forecast series, and renewable energy (such as photovoltaic and wind power) output forecast series for future forecast periods (e.g., the next 24 hours or longer). These forecast data can be generated using time series analysis, neural networks, and other forecasting methods, based on historical data combined with weather forecasts, holiday information, and other factors.

[0027] Secondly, a three-dimensional carbon potential energy surface is constructed. A three-dimensional mesh surface is built using time, power, and carbon intensity, and the carbon potential energy function value is calculated at each mesh point. The design of the carbon potential energy function can comprehensively consider grid carbon intensity, time-of-use pricing, and renewable energy consumption.

[0028] Then, a gradient descent algorithm is applied to search for carbon potential peaks and valleys on the three-dimensional carbon potential energy surface. Specifically, multiple search points are randomly initialized on the three-dimensional mesh surface, and the search points are iteratively moved along the negative gradient direction until convergence, where the convergence point represents the carbon potential energy valley (local minimum point); similarly, the search points are iteratively moved along the positive gradient direction until convergence, where the convergence point represents the carbon potential energy peak (local maximum point). This algorithm can automatically identify periods of low carbon potential energy valleys and periods of high carbon potential energy peaks. These identification results will serve as important inputs for subsequent hierarchical control.

[0029] This step visualizes and computes carbon constraints by mapping multidimensional time-varying parameters to a three-dimensional carbon potential energy surface, providing a clear potential difference-driven mechanism for the self-organized optimization of energy flow. It can intuitively identify periods with a high proportion of dirty electricity and periods with a high proportion of clean electricity in the grid, thereby proactively avoiding drawing electricity from the grid during high-carbon periods and fully utilizing clean electricity during low-carbon periods in subsequent scheduling.

[0030] S200: During the peak period of high carbon potential energy, a mixed integer linear programming model is used to generate the carbon economic baseline trajectory for the first cycle in the future, with the objective function being the minimum weighted sum of operating costs and carbon emission costs.

[0031] The main purpose of this step is to optimize energy consumption strategies during high-carbon periods from a global perspective, while ensuring the economic efficiency of system operation, and to generate a baseline operating trajectory for the first cycle in the future (e.g., the next 24 hours), providing a tracking target for intraday and real-time control.

[0032] The identified carbon potential peaks and valleys, predicted sequences, and system equipment parameters are input into a pre-built mixed-integer linear programming (MILP) model. The objective function of the MILP model is designed to minimize the weighted sum of operating costs and carbon emission costs. Operating costs include the cost of purchasing electricity from the grid and equipment operation and maintenance costs; carbon emission costs are calculated based on the amount of electricity purchased from the grid and the corresponding carbon intensity, and the weighting coefficients can be used to adjust the balance between economic efficiency and low carbon emissions.

[0033] By solving the MILP model described above, a carbon economic baseline trajectory for the first future cycle (e.g., 24 hours, which can be divided into 96 time periods, each lasting 15 minutes) is obtained. This baseline trajectory, while ensuring overall economic efficiency, has already planned an operational strategy to minimize grid power purchases during peak carbon energy periods.

[0034] This step utilizes the MILP model for day-ahead optimization, enabling a global approach to resource allocation over the next 24 hours and avoiding local optima. Simultaneously, incorporating carbon emission costs into the objective function ensures the baseline trajectory itself reflects a low-carbon orientation. The generated baseline trajectory provides a clear tracking target for subsequent intraday rolling optimization and real-time response, forming the foundation for multi-timescale control.

[0035] S300: Using the baseline trajectory as the target, perform a second-cycle rolling optimization through model predictive control, and output a correction control command. The duration of the second cycle is shorter than that of the first cycle.

[0036] The main purpose of this step is to address the forecasting deviations in renewable energy output and load by correcting the day-ahead baseline trajectory through rolling optimization, thereby improving the robustness and accuracy of control.

[0037] Using a 15-minute rolling period (second period), ultra-short-term revised forecast data for the next 4 hours (i.e., 16 time periods) is obtained, including updated grid carbon intensity forecasts, photovoltaic power output forecasts, electricity load forecasts, and heat load forecasts. A model predictive controller (MPC) is constructed with the baseline trajectory as the tracking target.

[0038] MPC, by tracking the weighted sum of the minimum deviation and the minimum rate of change of control, ensures that the system operates as close as possible to the globally optimal trajectory optimized in the previous day, while avoiding drastic changes in control actions, thus improving the stability of system operation.

[0039] At the beginning of each rolling cycle, the MPC controller obtains the measured state of charge of the energy storage at the current moment as the initial state and substitutes it into the model for monitoring. The above process is repeated in the next 15-minute cycle to achieve rolling optimization.

[0040] Predictions inevitably contain errors, especially for renewable energy sources such as solar power output. By employing intraday MPC rolling optimization, control commands can be monitored and executed promptly based on the latest measured data and ultra-short-term forecasts, ensuring the system always operates close to its optimal trajectory. Furthermore, the second cycle is shorter than the first cycle, guaranteeing both timely response and a balance between computational efficiency and control performance.

[0041] S400: Real-time monitoring of carbon intensity mutations triggers the deep reinforcement learning agent to output adjustment amounts superimposed on the correction instructions. Simultaneously, based on the carbon potential energy elasticity coefficient, it triggers an advanced collaborative mode to control the ground source heat pump to store heat in advance and discharge energy to replace grid power supply.

[0042] The main purpose of this step is to respond to emergencies such as sudden changes in grid carbon intensity by leveraging the rapid response capabilities of deep reinforcement learning and the cross-period adjustment capabilities of advanced collaborative models to achieve rapid carbon reduction.

[0043] At the real-time layer, changes in grid carbon intensity, renewable energy output, and load are monitored in real time with a 5-minute interval (or shorter event triggering). When the rate of change in carbon intensity exceeds a set threshold (e.g., 50 gCO2 / kWh per 5min), a pre-trained deep reinforcement learning (DQN) agent is triggered. This agent takes the current state as input, outputs a power adjustment, and adds it to the issued correction control command to generate the final execution command.

[0044] The agent can be trained offline based on historical operating data, using experience replay and target network mechanisms, enabling it to learn the ability to make optimal decisions quickly when carbon intensity changes abruptly.

[0045] Simultaneously, this step introduces the carbon potential elasticity coefficient as a quantitative indicator for triggering the advanced collaborative mode. The carbon potential elasticity coefficient is jointly determined by the energy storage state of charge, the current energy efficiency ratio of the ground source heat pump, and the real-time carbon intensity of the power grid, and is used to quantify the potential for reducing carbon emissions at the current moment through the coordinated scheduling of energy storage and ground source heat pumps.

[0046] By controlling the ground source heat pump to operate at its maximum energy efficiency ratio, the generated heat is stored in a heat storage tank and the storage time is set in advance (e.g., 2 hours). The heat demand during high-carbon periods is transferred to the current period in advance through heat energy storage. The energy storage system is controlled to discharge at a set discharge rate (e.g., 0.5C) to prioritize meeting load demand, reduce the purchase of electricity from the grid, and directly reduce carbon emissions during the current period. At the same time, the photovoltaic output is prioritized to directly supply the load, and any surplus is stored in the battery.

[0047] In addition, it should be noted that during the execution of the advanced cooperative mode, cooperative suppression can be performed on other control layers: the weight of tracking the baseline trajectory in the MPC objective function is temporarily reduced, and the action space of the DQN agent is limited to a preset safety range to avoid conflicts between multi-layer control instructions.

[0048] When the real-time carbon intensity of the power grid drops below the fifth threshold (e.g., 500 gCO2 / kWh) or the thermal storage tank is full, the advanced coordination mode is exited and normal control commands are restored.

[0049] Sudden and unpredictable changes in grid carbon intensity can lead to rapid decision-making by learning from historical data and coping experiences. Meanwhile, the proactive collaborative model fully leverages the slow-dynamic characteristics of the thermal system (heat load can be delayed and thermal storage tanks can store heat) and the energy time-shifting capabilities of energy storage. Through cross-period collaboration, energy demand during high-carbon periods is shifted to low-carbon periods in advance, reducing carbon dioxide emissions.

[0050] As a preferred embodiment of the present invention, constructing a three-dimensional carbon potential energy surface and identifying carbon potential peaks and valleys includes: A three-dimensional mesh surface was constructed with time as the horizontal axis, power as the vertical axis, and carbon intensity as the vertical axis. The carbon potential energy function value was calculated at each mesh point. The gradient descent algorithm is used to search for local minima and local maxima on the surface to determine the peaks and valleys of the carbon potential. The low carbon potential valley period and the high carbon potential peak period are determined based on the carbon potential peak and valley.

[0051] First, we obtain the grid carbon intensity forecast series, time-of-use electricity price forecast series, and renewable energy (such as photovoltaic and wind power) output forecast series for future forecast periods (e.g., the next 24 hours or longer). These forecast data can be generated using time series analysis, Long Short-Term Memory (LSTM) networks, and other forecasting methods, based on historical data combined with weather forecasts and holiday information. To construct the three-dimensional carbon potential energy surface, the three dimensions of time, power, and carbon intensity need to be discretized.

[0052] Specifically, the timeline is divided into N equally spaced time periods, for example, with 15-minute intervals, the next 24 hours are divided into 96 time periods, each corresponding to a specific time point. (i=1,2,…,N).

[0053] The power axis is discretized into M power levels, covering the possible range of grid power extraction for the system, such as from the minimum power extraction to the maximum allowable power extraction, divided into several discrete values. (j=1,2,…,M).

[0054] The carbon intensity axis is divided into L carbon intensity levels based on the predicted carbon intensity range, with each level corresponding to a carbon intensity value. (k=1,2,…,L).

[0055] This constructs an N×M×L three-dimensional mesh surface, where each mesh point corresponds to a specific combination of time, power, and carbon intensity.

[0056] At each grid point ( , , ) calculate the carbon potential energy function value Φ( , , This function reflects the overall carbon potential at a given moment and power consumption level, resulting in a higher carbon potential value during periods of high carbon intensity, high electricity prices, and insufficient renewable energy output, and a lower carbon potential value during periods of low carbon intensity, low electricity prices, and abundant renewable energy.

[0057] , in: This represents the carbon intensity value (unit: gCO2 / kWh) corresponding to the kth carbon intensity level. for Time-of-use electricity price at any given moment (unit: yuan / kWh); for Forecasted renewable energy output at any given time (unit: kW); The grid power draw (unit: kW) corresponds to the j-th power level. This indicates a positive value, meaning that when the grid's power input exceeds the renewable energy output, it means that additional electricity needs to be purchased from the grid, contributing a positive carbon potential; conversely, if renewable energy has already met the demand, this item is zero. These are weighting coefficients, which can be calibrated using historical data or adjusted according to actual operational needs to meet [the requirements]. ,and To ensure that carbon intensity is the core driving factor.

[0058] It should be noted that carbon intensity can also be used as the primary variable, with electricity price and renewable energy as auxiliary factors, for example, by using a weighted summation. The Φ value calculated for each grid point constitutes a three-dimensional numerical matrix, reflecting the distribution of carbon potential energy in the time-power-carbon intensity space, and there are no restrictions on this.

[0059] On the constructed 3D mesh surface, the gradient descent algorithm is applied to search for local minima (carbon potential valleys) and local maxima (carbon potential peaks). Multiple initial points are randomly generated on the 3D mesh surface, each corresponding to a mesh coordinate (t, P, C). To ensure comprehensive search coverage, uniform random sampling or a multi-starting-point strategy can be adopted.

[0060] For each initial point, calculate the gradient direction in its neighborhood. The gradient can be approximated by the difference of function values ​​between adjacent grid points, for example: .

[0061] Iteratively move along the negative gradient direction, updating the coordinates of the point, until the convergence condition is met (such as the gradient magnitude being less than a preset threshold or the step size being less than a threshold). The convergence point is a local minimum point, representing the carbon potential valley of the region.

[0062] Iteratively move along the positive gradient direction until convergence. The convergence point is a local maximum, representing the carbon potential energy peak in that region. By searching from multiple starting points, multiple local minima and maxima can be obtained.

[0063] In one embodiment, the low carbon potential valley period and the high carbon potential peak period are determined based on the carbon potential peaks and valleys, specifically as follows: A first threshold is determined based on multiple local minima of the surface, and a second threshold is determined based on multiple local maxima of the surface. Identify low carbon potential energy valley periods when carbon intensity is below the first threshold and high carbon potential energy peak periods when carbon intensity is above the second threshold.

[0064] Collect all local minima found, extract their corresponding carbon intensity values, and take the average or a certain quantile of the carbon intensity of these minima as the first threshold (e.g., take the minimum, average, or lower quantile of the carbon intensity of all minima). Similarly, extract the carbon intensity values ​​from the local maxima to determine the second threshold.

[0065] Iterate through all time points and identify time periods that meet the following conditions: Low-carbon potential valley period, the period when the predicted value of the grid carbon intensity is lower than the first threshold; High carbon potential peak period, the period when the predicted value of grid carbon intensity is higher than the second threshold.

[0066] These identification results will serve as important inputs for subsequent hierarchical control, guiding the direction of energy flow: during peak carbon energy periods, the system should proactively utilize local green electricity and energy storage to reduce electricity purchases from the grid; during valley periods of low carbon energy, it can appropriately purchase electricity from the grid, or even charge energy storage, to reserve energy for subsequent high carbon periods.

[0067] As a preferred embodiment of the present invention, the carbon economy benchmark trajectory includes at least the energy storage charging and discharging benchmark power, the ground source heat pump benchmark output, and the grid power purchase benchmark power.

[0068] Based on the identified carbon potential peaks and valleys, the system takes a holistic approach, coordinating the operational strategies of all equipment within the first cycle (e.g., the next 24 hours) to generate a baseline operating trajectory that balances economic efficiency and low carbon emissions. This baseline trajectory provides a clear tracking target for subsequent intraday rolling optimization and real-time response, ensuring that while pursuing global optimization, the system can proactively utilize local green electricity and energy storage resources during periods of high carbon potential peaks, thereby reducing carbon dioxide emissions at the source.

[0069] Specifically, the carbon economy benchmark trajectory includes at least three core components: benchmark power for energy storage charging and discharging, benchmark output for ground source heat pumps, and benchmark power for grid purchases. These three benchmark quantities are interconnected and work together to form the park's complete energy consumption plan for the first cycle in the future.

[0070] First, a mixed-integer linear programming model is constructed. The constraints of the mixed-integer linear programming model include power balance, heat power balance, energy storage system state constraints, ground source heat pump operation constraints, and power grid interaction constraints. The carbon economy baseline trajectory is obtained based on grid carbon intensity, time-of-use electricity prices, and renewable energy output forecasts, combined with system equipment parameters.

[0071] The equipment parameters include: the rated capacity, maximum charge and discharge power, charge and discharge efficiency, and upper and lower limits of state of charge of the energy storage system; the rated power, adjustable range of energy efficiency ratio (COP) (e.g., 2.8-4.2), and start-stop constraints of the ground source heat pump; the maximum power tracking range of the photovoltaic inverter; and the upper limit of the power purchased and sold with the grid.

[0072] Secondly, define the objective function. The objective function is to minimize the weighted sum of operating costs and carbon emission costs, expressed as: , Where T is the number of time periods in the first cycle (e.g., 24 hours is divided into 96 15-minute time periods); Price(t) is the time-of-use electricity price for time period t; Let t be the power purchase capacity of the power grid during time period t (the variable to be optimized); C(t) is the predicted value of the carbon intensity of the power grid during time period t; λ is the carbon emission cost weighting coefficient, used to balance economic efficiency and low carbon emissions. Let t be the operating and maintenance cost of equipment u (including energy storage, ground source heat pumps, etc.) during time period t. The objective function is designed so that the system tends to purchase electricity from the grid during low-carbon energy valley periods and tends to reduce electricity purchases during high-carbon energy peak periods.

[0073] The constraints of the mixed-integer linear programming model cover multiple aspects. One constraint is the power balance constraint: Grid power purchase + Photovoltaic output + Energy storage discharge power = Electric load + Energy storage charging power + Ground source heat pump power consumption. This constraint ensures that power supply and demand are balanced at any given time.

[0074] Thermal power balance constraint: Ground source heat pump heat generation power + thermal storage tank heat release power = heat load + thermal storage tank charging power. This constraint ensures a balance between heat supply and demand.

[0075] Energy storage system state constraints include the dynamic recursive equation for the state of charge (POC), upper and lower limits for the POC, upper and lower limits for charge / discharge power, and mutual exclusion constraints for charge / discharge states. The POC recursive relationship can be expressed as: ， Where SOC(t) is the state of charge at the end of time period t; , These are charging efficiency and discharging efficiency, respectively. , These are the charging power and discharging power, respectively, and they satisfy... · =0 (mutual exclusion constraint); ≤SOC(t)≤ .

[0076] Ground source heat pump operating constraints include the energy efficiency ratio (EER) relationship between power consumption and heat production, the constraint on the continuously adjustable range of the EER, and start-stop state constraints. The EER relationship can be expressed as: , in, For the power consumption of ground source heat pumps, The heat output is represented by COP(t), which is the energy efficiency ratio and can be continuously adjusted within a preset range (e.g., 2.8-4.2).

[0077] Interaction constraints with the power grid: Maximum power purchase limit constraint, i.e., 0 ≤ ≤ Ensure that the grid connection capacity is not exceeded.

[0078] Other equipment operating constraints include power constraints for photovoltaic inverters and capacity constraints for thermal storage tanks.

[0079] Finally, the model is solved and a baseline trajectory is output. Commercial solvers (such as Gurobi and CPLEX) are used to solve the MILP model to obtain the optimization results for each period within the first future cycle. The baseline power for energy storage charging and discharging, the baseline output of the ground source heat pump, and the baseline power purchased by the grid are used as core outputs to constitute the carbon economy baseline trajectory.

[0080] The reference power for energy storage charging and discharging, the reference output of ground source heat pumps, and the reference power for grid-purchased electricity are closely coupled through electrical power balance constraints, thermal power balance constraints, and carbon emission targets, together forming a complete carbon-economic optimal energy use scheme. Energy storage is responsible for the time shifting of electricity, ground source heat pumps are responsible for the electrothermal coupling conversion, and grid-purchased electricity serves as the interface with external systems. The synergistic optimization of these three components can meet the demand for electrical and thermal loads while minimizing the weighted sum of operating costs and carbon emission costs.

[0081] In a preferred embodiment of the present invention, the second-cycle rolling optimization is performed using model predictive control, specifically as follows: A model predictive controller is constructed using the aforementioned carbon economy baseline trajectory as the tracking target; Real-time tracking deviation and control rate of change are obtained, and the unfavorable baseline value is obtained. When the baseline unfavorable value is greater than the third threshold, the system is reprogrammed using a mixed-integer linear programming model to obtain a corrected control command.

[0082] The main objective of this implementation method is to improve the robustness of the control system to renewable energy fluctuations and load forecasting errors by using online judgment of the generated carbon economy baseline trajectory through rolling optimization to address the deviation between day-ahead forecasts and actual operation. Model predictive control (MPC), with its ability to explicitly handle system constraints, predict future system dynamics, and perform rolling optimization, is widely used in complex industrial process control. Introducing MPC into the inner layer of the multi-energy collaborative control system in the industrial park, with a 15-minute rolling cycle, tracks the day-ahead baseline trajectory while also considering the smoothness of control changes, forming a collaborative architecture of global planning + local correction.

[0083] First, a model predictive controller (MPC) is constructed using the output carbon economy baseline trajectory as the tracking target. This controller, based on the system's state-space model, predicts the system dynamics over a finite time domain. At each sampling moment, the MPC solves a finite-time open-loop optimization problem based on the currently measured system state (including energy storage state of charge, thermal storage tank state, current actual load, and renewable energy output), combined with ultra-short-term forecast information (grid carbon intensity, photovoltaic output, electrical load, and thermal load for the next 4 hours).

[0084] The prediction model can be represented in discrete-time state-space form: , in, It is a state vector, including the energy storage state of charge, the heat storage tank's heat storage capacity, etc. The control input vector includes energy storage charging and discharging power, ground source heat pump power consumption, etc. These are measurable disturbance vectors, including photovoltaic output and load demand; , , This is the system matrix.

[0085] The objective function of the model predictive controller is designed as a weighted sum of the minimum tracking deviation and the minimum rate of change of control to construct the baseline adverse value, specifically expressed as: , in, To predict the length of the time domain (e.g., 16 future time periods, corresponding to 4 hours); for Power purchased by the grid at any given time (optimization variable); This is the generated grid power purchase benchmark. Power consumption of the ground source heat pump (optimization variable); This serves as the reference output for the ground source heat pump. For the vector of rates of change of control variables; , , This is a weight matrix used to adjust the balance between different objectives.

[0086] The objectives are: first, to ensure that the control trajectory is as close as possible to the baseline trajectory optimized previously, maintaining overall economic efficiency and a low-carbon orientation; and third, to penalize drastic changes in control variables, avoiding frequent equipment start-ups and shutdowns or large power fluctuations, extending equipment lifespan, and ensuring stable system operation.

[0087] This value quantifies the degree to which the current optimized trajectory deviates from the baseline trajectory and the severity of control changes. When the unfavorable baseline value exceeds a preset third threshold... When this occurs, it indicates that due to the accumulation of prediction bias or increased external disturbances, the current baseline trajectory is no longer applicable, and continued tracking may cause the system to deviate from its optimal state.

[0088] At this point, the system automatically triggers a replanning mechanism, pausing the regular rolling optimization of intraday MPC. It then feeds back the current measured state and the latest ultra-short-term forecast data to the mixed-integer linear programming model for a new global optimization solution, generating an updated baseline trajectory. When the intraday deviation exceeds the tolerance range, a new globally optimal baseline is obtained by rerunning the MILP model, and then MPC continues tracking.

[0089] When replanning is triggered, the MILP model uses the solution from the previous baseline trajectory as the initial feasible solution, employing warm-start technology to accelerate the solution process and ensure that the updated baseline trajectory is obtained within an acceptable timeframe. Through MPC rolling optimization, control commands can be promptly corrected based on the latest measured data and ultra-short-term predictions, ensuring the system always operates close to the optimal trajectory. The 15-minute rolling cycle guarantees both timely response and sufficient computation time to handle optimization problems, achieving a balance between computational efficiency and control performance.

[0090] In a preferred embodiment of the present invention, real-time monitoring of sudden changes in carbon intensity triggers the deep reinforcement learning agent to output an adjustment amount superimposed on the correction instruction, specifically as follows: Real-time monitoring of sudden changes in the carbon intensity of the power grid; when the rate of change in carbon intensity exceeds a set threshold, a deep reinforcement learning agent is triggered.

[0091] Using the current tracking deviation and correction control command as input, the output adjustment amount is superimposed on the correction control command to generate the final execution command.

[0092] To address sudden events such as abrupt changes in grid carbon intensity, deep reinforcement learning's rapid response capability allows for fine-tuning of intraday layer output control commands within seconds. This enables the system to respond promptly to drastic fluctuations in carbon intensity, preventing unnecessary carbon emissions during high-carbon periods due to response delays. Deep reinforcement learning, through neural network approximation functions, can handle decision-making problems in high-dimensional state spaces, and its experience replay and target network mechanisms effectively improve training stability.

[0093] First, monitor changes in the grid carbon intensity factor in real time with a 5-minute interval (or a shorter event triggering method). Specifically, set a sliding window to calculate the rate of change of carbon intensity at the current moment compared to the previous moment (or the average of several previous moments): , in, Let t be the real-time carbon intensity of the power grid. The monitoring period is set at 5 minutes. When the rate of change in carbon intensity is detected... Exceeding the preset threshold When the carbon intensity reaches 50 gCO2 / kWh per 5 min (for example), the system determines that a sudden carbon intensity event has occurred and immediately triggers a deep reinforcement learning agent to make a decision response.

[0094] Under normal circumstances, the system operates according to the correction instructions output by the MPC within the day to ensure the stability of control; only when the carbon intensity fluctuates drastically will the real-time response layer be activated to intervene, which ensures the efficiency of control and avoids the computational burden caused by frequent calls from the agent.

[0095] After triggering a deep reinforcement learning agent, its input state needs to be constructed first. This state space should comprehensively reflect the current operating status of the system and its deviation from the baseline trajectory, specifically including the following dimensions: Carbon strength deviation, the deviation between real-time carbon strength and the generated reference carbon strength. This variable directly reflects the degree of deviation of the current carbon intensity from the current planning, and is the core basis for decision-making.

[0096] The energy storage state of charge (SOC(t)) is the current state of charge of the stored energy, used to determine whether the energy storage has sufficient discharge capacity to cope with high carbon periods.

[0097] Deviation in renewable energy output, specifically the discrepancy between actual photovoltaic output and ultra-short-term revised forecasts. This variable reflects the actual fluctuations in renewable energy.

[0098] Load deviation, the deviation between actual load and ultra-short-term correction forecast. It is used to determine changes in demand on the load side.

[0099] Current-moment correction control commands, including energy storage charging and discharging power commands. and ground source heat pump power command This serves as a benchmark value for the agent's adjustments.

[0100] The information from the above five dimensions is concatenated into a state vector. This state space serves as the input to the deep reinforcement learning agent. This state space design ensures that the agent can fully perceive the system's operating state, prediction deviations, and current control commands, providing sufficient basis for making reasonable adjustment decisions. The state space design is the foundation for DQN's ability to approximate the Q-value function, mapping high-dimensional states to action values ​​through neural networks.

[0101] The action space is defined as the discrete set of values ​​for the energy storage discharge power adjustment and the ground source heat pump power adjustment. Considering the feasibility of engineering implementation and the stability of control, the adjustment values ​​are discretized: Energy storage discharge power adjustment This corresponds to different discharge rate adjustments (e.g., fine-tuning within the 0.2C, 0.5C, 0.8C, 1.0C, etc.) range.

[0102] Ground source heat pump power adjustment This corresponds to a fine-tuning of the operating point of the ground source heat pump's coefficient of performance (COP) (e.g., adjustment within the range of 2.8-4.2).

[0103] The combination of two adjustment values ​​constitutes the motion space. There are 25 discrete action options in total. This discretization design ensures that the agent has a certain degree of adjustment flexibility, while avoiding the exploration difficulties and training instability problems caused by continuous action space.

[0104] The reward function is crucial for guiding a deep reinforcement learning agent to learn the optimal strategy. The reward function designed in this invention comprises two parts: a carbon emission change penalty and a regulation cost penalty, specifically expressed as follows: , in, Due to the implementation of adjustment actions The resulting change in carbon emissions can be approximated by multiplying the change in the power purchased by the grid by the current carbon intensity; The L1 norm of the adjustment action is the sum of the absolute values ​​of the energy storage adjustment and the heat pump adjustment. It is a balance coefficient used to adjust the weight between carbon emission reduction and the adjustment cost.

[0105] This reward function encourages agents to achieve the greatest carbon emission reduction effect with the least adjustment cost (i.e., the smallest possible power adjustment). When the agent's actions effectively reduce carbon emissions, A negative value (reduced carbon emissions) results in a larger overall reward; conversely, if the action leads to an increase in carbon emissions or the adjustment is too large, the reward will decrease. This reward shaping method breaks down sparse carbon reduction targets into phased rewards, which helps accelerate convergence.

[0106] As a preferred embodiment of the present invention, the advanced cooperative mode is triggered based on the carbon potential energy elastic coefficient, specifically as follows: The carbon potential energy elasticity coefficient is calculated in real time. This coefficient is jointly determined by the energy storage state of charge, the current energy efficiency ratio of the ground source heat pump, and the real-time carbon intensity of the power grid. When the real-time carbon intensity of the power grid exceeds the fourth threshold and the carbon potential elasticity coefficient exceeds the fifth threshold, the advanced coordination mode is triggered.

[0107] During high-carbon periods (when grid carbon intensity exceeds a certain threshold) and when the system possesses sufficient potential for coordinated carbon reduction, a joint operation mode of proactively triggering ground-source heat pumps for advance heat storage and energy storage discharge to replace grid power supply is initiated. This shifts energy demand during high-carbon periods to low-carbon periods through thermal energy storage, achieving coordinated carbon reduction across time periods and energy types. A carbon potential elasticity coefficient is constructed to quantify the coordinated scheduling potential of energy storage and heat pumps. Based on the combination of this coefficient and the carbon intensity threshold, precise triggering of the advanced coordinated mode is achieved.

[0108] The carbon potential elasticity coefficient is determined by three core factors: the state of charge of energy storage, the current energy efficiency ratio of the ground source heat pump, and the real-time carbon intensity of the power grid. An exemplary calculation formula is as follows: , in, Let t be the state of charge of the energy storage system. This represents the upper limit of the energy storage state of charge. Let be the energy efficiency ratio of the ground source heat pump at time t. and These are the adjustable upper and lower limits of the energy efficiency ratio of the ground source heat pump (e.g., 4.2 and 2.8). Let t be the real-time carbon intensity of the power grid. This is the carbon intensity threshold (i.e., the minimum carbon intensity requirement to trigger the advanced synergistic mode). and These represent the historical maximum and minimum carbon intensity values, respectively. , , For the weighting coefficients, satisfying This is used to adjust the relative importance of the three factors in the assessment of synergistic potential, and can be calibrated based on system characteristics or operational experience (e.g., , , ).

[0109] First item It characterizes the dispatchability of energy storage. The higher the state of charge of energy storage, the greater the available discharge capacity, the stronger the ability to replace grid power supply during high-carbon periods, and the greater the potential for synergy.

[0110] Second item This characterizes the potential for improving the energy efficiency of heat pumps. The lower the current energy efficiency ratio, the greater the potential to improve it to the high-efficiency range through optimized operation, resulting in more heat generated per unit of electricity consumption and a more significant synergistic carbon reduction effect.

[0111] Third item This characterizes the degree to which carbon intensity exceeds the threshold. The greater the current carbon intensity exceeds the threshold, the greater the urgency and necessity for taking coordinated measures to reduce carbon emissions.

[0112] The real-time carbon intensity of the power grid exceeds the fourth threshold. This fourth threshold (e.g., 550 gCO2 / kWh) is set based on the identification of high carbon potential peaks in carbon potential field modeling. This threshold indicates that the system has entered or is about to enter a high-carbon period, at which point purchasing electricity from the grid will generate higher range-bound carbon emissions, requiring proactive use of local resources for substitution. The carbon intensity threshold can be dynamically adjusted based on factors such as the grid characteristics of the area where the industrial park is located and carbon reduction targets.

[0113] The carbon potential elasticity coefficient exceeds the fifth threshold. This fifth threshold (e.g., 0.6) is used to determine whether the system possesses sufficient synergistic carbon reduction potential. When the carbon potential elasticity coefficient exceeds this threshold, it indicates that the current energy storage state of charge is sufficient, there is significant room for improvement in heat pump efficiency, and the carbon intensity exceeds the limit severely. The combined effect of these three factors makes carbon reduction through energy storage discharge and heat pump heat storage highly feasible and effective. The fifth threshold can be calibrated based on historical operating data or simulation analysis, and can also be dynamically adjusted according to factors such as season and weather.

[0114] If triggered solely by carbon intensity, the collaborative mode might be forcibly activated when the energy storage SOC is low or the heat pump is already running efficiently, resulting in poor regulation or even affecting system stability. If triggered solely by the elasticity coefficient, resources might be prematurely deployed before carbon intensity exceeds the limit, causing unnecessary energy loss. The dual-threshold logic ensures that the proactive collaborative mode is activated at the necessary and appropriate time, achieving optimal resource allocation.

[0115] The second invention provides a computing device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the method.

[0116] Therefore, it is possible to achieve any effect in the carbon-constrained multi-energy coordinated integrated energy supply control method for industrial parks, which will not be elaborated here.

[0117] For any parts not mentioned in this invention, existing technologies can be used or referenced.

[0118] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

[0119] The above description is merely an embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of the present invention should be included within the scope of the claims of the present invention.

Claims

1. A carbon-constrained multi-energy coordinated integrated energy supply control method for industrial parks, characterized in that, include: Based on the prediction of grid carbon intensity, time-of-use electricity price and renewable energy output, a three-dimensional carbon potential energy surface is constructed to identify carbon potential peaks and valleys and determine the low carbon potential energy valley period and the high carbon potential energy peak period. During the peak period of high carbon potential energy, a mixed integer linear programming model is used to generate the carbon economic baseline trajectory for the first future cycle with the objective function of minimizing the weighted sum of operating costs and carbon emission costs. Using the baseline trajectory as the target, a second-cycle rolling optimization is performed through model predictive control, and a correction control command is output. The duration of the second cycle is shorter than that of the first cycle. Real-time monitoring of sudden changes in carbon intensity triggers the deep reinforcement learning agent to output adjustment amounts superimposed on the correction instructions. Simultaneously, based on the carbon potential energy elasticity coefficient, an advanced collaborative mode is triggered to control the ground source heat pump to store heat in advance and discharge energy to replace grid power supply.

2. The method according to claim 1, characterized in that, Constructing a three-dimensional carbon potential energy surface and identifying carbon potential peaks and valleys includes: A three-dimensional mesh surface was constructed with time as the horizontal axis, power as the vertical axis, and carbon intensity as the vertical axis. The carbon potential energy function value was calculated at each mesh point. The gradient descent algorithm is used to search for local minima and local maxima on the surface to determine the peaks and valleys of the carbon potential. The low carbon potential valley period and the high carbon potential peak period are determined based on the carbon potential peak and valley.

3. The method according to claim 2, characterized in that, Based on the aforementioned carbon potential peaks and valleys, the periods of low-carbon potential valleys and high-carbon potential peaks are determined as follows: A first threshold is determined based on multiple local minima of the surface, and a second threshold is determined based on multiple local maxima of the surface. Identify low carbon potential energy valley periods when carbon intensity is below the first threshold and high carbon potential energy peak periods when carbon intensity is above the second threshold.

4. The method according to claim 1, characterized in that, The carbon economy benchmark trajectory includes at least the energy storage charging and discharging benchmark power, the ground source heat pump benchmark output, and the grid power purchase benchmark power.

5. The method according to claim 4, characterized in that, The constraints of the mixed-integer linear programming model include power balance, heat balance, energy storage system state constraints, ground source heat pump operation constraints, and power grid interaction constraints. The carbon economy baseline trajectory is obtained based on grid carbon intensity, time-of-use electricity prices, and renewable energy output forecasts, combined with system equipment parameters.

6. The method according to claim 1, characterized in that, The second-cycle rolling optimization is performed using model predictive control, specifically as follows: A model predictive controller is constructed using the aforementioned carbon economy baseline trajectory as the tracking target; Real-time tracking deviation and control rate of change are obtained, and the unfavorable baseline value is obtained. When the baseline unfavorable value is greater than the third threshold, the system is reprogrammed using a mixed-integer linear programming model to obtain a corrected control command.

7. The method according to claim 1, characterized in that, Real-time monitoring of sudden changes in carbon intensity triggers the deep reinforcement learning agent to output an adjustment amount that is superimposed on the correction instruction, specifically as follows: Real-time monitoring of sudden changes in the carbon intensity of the power grid; when the rate of change in carbon intensity exceeds a set threshold, a deep reinforcement learning agent is triggered.

8. The method according to claim 7, characterized in that, Triggering a deep reinforcement learning agent, specifically: Using the current tracking deviation and correction control command as input, the output adjustment amount is superimposed on the correction control command to generate the final execution command.

9. The method according to claim 1, characterized in that, The advanced collaborative mode is triggered based on the carbon potential energy elasticity coefficient, specifically as follows: The carbon potential energy elasticity coefficient is calculated in real time. This coefficient is jointly determined by the energy storage state of charge, the current energy efficiency ratio of the ground source heat pump, and the real-time carbon intensity of the power grid. When the real-time carbon intensity of the power grid exceeds the fourth threshold and the carbon potential elasticity coefficient exceeds the fifth threshold, the advanced coordination mode is triggered.

10. A computing device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method as described in any one of claims 1 to 9.

Images