Peak regulation and frequency modulation heat supply method for multi-element energy storage coupling coal power unit

Abstract

The invention relates to the technical field of power supply, and particularly discloses a peak and frequency regulation heat supply method for a multi-element energy storage coupling coal power unit, which is used for solving the problems that only active frequency is optimized, electric heating coordination and charge state constraint are lacked and time delay compensation is insufficient in the prior art. The method provided by the invention comprises the steps of system modeling, controller design, coupling constraint, adaptive parameter adjustment, delay compensation and distributed consensus. According to the method, a dual-domain electric power-thermal interconnection model is constructed, a cooperative controller based on delay compensation distributed model prediction control is designed, electric-thermal cooperative penalty and charge state constraint are introduced into a target function, a prediction window and weight parameters are adaptively adjusted, and network delay is identified and compensated online. And a consensus algorithm is adopted to realize multi-region power-heat collaborative optimization.

Description

Technical Field

[0001] The present invention relates to the field of power supply technology, and in particular to a peak-shaving and frequency-regulating heating method for a multi-element energy storage coupled coal-fired power unit. Background Art

[0002] With the large-scale grid connection of renewable energy sources such as wind power and photovoltaics, the frequency fluctuations of the power system have intensified, placing higher demands on frequency regulation resources. Although traditional thermal power units have advantages such as large rotational inertia and wide frequency regulation range, their response speed and regulation accuracy are relatively limited, making it difficult to meet the modern frequency regulation requirements of "fast regulation, precision, and controllability". Energy storage systems can achieve charge and discharge switching in milliseconds and have strong bidirectional regulation capabilities. They are considered an important means to improve the dynamic performance of the power grid. The open CNKI document "Research on Wide-Area Multi-Energy Storage Coordinated Control Based on Distributed Model Predictive Control" proposed for the first time in a two-region interconnected power system the coordinated control of distributed battery energy storage and pumped storage based on the DC-DMPC (Delay Compensation Distributed Model Predictive Control) algorithm framework. A two-region LFC mathematical model containing battery energy storage (BESS) and pumped storage units (PSP) was constructed, a distributed MPC (Model Predictive Control) controller was designed, and ADMM (Alternating Direction Method of The MPC objective function uses the Multipliers consensus algorithm to achieve inter-unit power coordination. A simple delay compensation strategy is proposed for fixed communication delays, and simulation verification is completed in MATLAB under typical operating conditions. Although phased results have been achieved in framework design and algorithm verification, this only optimizes active frequency regulation and does not incorporate unit waste heat or heat network review into the model, which is not conducive to the intensive utilization of combined power and heat resources. Furthermore, the SOC constraints of batteries and pumped-storage units are not included in the MPC objective function, which can lead to unit state of charge drift during long-term operation. The controller prediction window and weights are fixed, and the MPC window length and weighting coefficients cannot be adaptively adjusted based on renewable energy power fluctuations and sudden load changes. Delay compensation only uses fixed delay compensation, and there is a lack of online estimation and dynamic adjustment mechanisms for network jitter, packet loss, and delay fluctuations, resulting in a large gap between the simulation environment and actual deployment. Summary of the Invention

[0003] The technical problem to be solved by the present invention is to provide a peak-shaving and frequency-regulating heating method for multi-source energy storage coupled coal-fired power units. By introducing electric-thermal coupling and SOC constraints into distributed model predictive control, supplemented by multi-condition adaptive prediction windows and weight adjustment, and online delay identification, the ADMM consensus algorithm is used to achieve the optimal electric-thermal synergy peak-shaving and frequency regulation of multi-source energy storage and coal-fired power units under the DC-DMPC framework.

[0004] To achieve the above object, the present invention provides the following technical solutions: A peak-shaving and frequency-regulating heating method for a multi-element energy storage coupled coal-fired power unit comprises the following steps: A. Build a two-region interconnected power-thermal system model that includes battery energy storage units, pumped storage units, and heating network loads; B. Design a distributed model predictive controller based on DC-DMPC, with frequency deviation, power balance, energy storage unit SOC, and heating network temperature deviation as optimization objectives; C. Introducing the electrical-thermal coupling term and the energy storage unit SOC constraint into the MPC objective function; D. The controller adaptively adjusts the prediction window length and weighting coefficient based on renewable energy output and load fluctuations; E. Adopt a dynamic delay identification strategy for network jitter and packet loss, and embed a delay compensation model into the prediction model; F. Each regional controller coordinates the power and heat between units through the ADMM consensus algorithm and uses the alternating direction multiplier method to decompose and solve.

[0005] To significantly shorten the system's frequency response time under renewable energy disturbances, improve the discharge / charge efficiency of energy storage units, accurately reproduce the spatiotemporal dynamic processes of the hydro-electromechanical and thermal networks, and ensure that system frequency stability and thermal network temperature deviations are effectively controlled under large-scale grid connection and sudden heat load conditions: As a further solution of the present invention, in step A, based on the classic first-order RC circuit SOC-voltage model, combined with the field-calibrated nonlinear efficiency curve and temperature-dependent internal resistance, the temperature, SOC and power coupling are dynamically introduced into the state equation, the discharge / charge loss and response speed under different working conditions are adaptively optimized, and a capacity decay model based on the cumulative number of cycles and depth is incorporated, and the decay rate is set as the decision variable; in order to take into account the millisecond-level frequency regulation and minute-level energy scheduling requirements, a hybrid time domain model is constructed using dual discrete step lengths of 0.1s and 60s, and a dual-mode state equation is designed for the two modes of pumping and power generation, with an embedded hydraulic-electromechanical coupling term containing water head changes. The first-order lag pipeline water hammer approximation term is used to capture pressure oscillations and power jitters when flow rates suddenly change. At the same time, the effective head and water output capacity are dynamically calculated based on the height difference between the two areas and the pipeline length, and coordinated with the grid demand. In terms of the heating network, the regional heating network is divided into multi-node pipe sections to establish heat capacity and thermal resistance equations, simulate the temporal and spatial diffusion of pipeline heat loss, flow rate and supply and return water temperature difference, and allow MPC to directly adjust the power output of battery energy storage units and pumped storage units in the dynamic mapping model of heat pump units to meet the temperature and heat requirements of the heating network. Combined with local meteorological and historical heat consumption curves, the recursive least squares method is used to update the load forecasting model online and quickly correct sudden heating demands.

[0006] A two-region interconnected power-thermal coupling scenario where wind power, photovoltaic bases, and regional thermal load centers are decentralized and connected through a dual energy storage system of battery energy storage and pumped hydropower storage. This scenario requires simultaneous assurance of system frequency stability, energy storage safety, and thermal network temperature balance during severe cold weather or sharp load fluctuations, while also facing challenges such as network jitter and packet loss.

[0007] As a further solution of the present invention, step B includes but is not limited to the following specific working conditions: Stable operating condition: A stable operating condition is identified only if the system frequency standard deviation is less than 0.02 Hz, the power deviation standard deviation is less than 5% of the rated power, and the communication delay jitter is consistently less than 50 ms for five consecutive minutes. In this case, the prediction window is fixed at 10 minutes. The controller optimizes the frequency, power, energy storage unit SOC, and heating network temperature deviation using established empirical ratios. The ADMM consensus iteration uses a "lightweight" mode with a relaxation factor of 1 and five iterations. Rapid Fluctuation Conditions: If the sudden change rate of load or renewable energy output exceeds 20% of the rated power within any minute, or the absolute value of the frequency slope is greater than 0.05 Hz / s, it is determined to be a rapid fluctuation condition. In this case, the prediction window is shortened to 5 minutes, the frequency deviation weight is increased to 2, and the remaining weights are adjusted proportionally. The ADMM consensus algorithm automatically increases the relaxation factor from 1 to 5 and the number of iterations to 8 based on the local residual. If the local MPC solution exceeds 0.2 seconds, the last available solution is automatically used to ensure real-time response. Heating network peak load condition: When the return water temperature deviation exceeds 3°C for three consecutive minutes and the heat load growth rate exceeds 10% of the rated heat within five minutes, it is identified as a heating network peak load condition. At this time, the heating network temperature deviation weight is increased to 3, and the frequency deviation weight is reduced to 0.5. A dual-channel recursive least squares forecast based on weather forecasts and historical heat consumption curves is run in parallel, and a heat lower limit constraint is added to the MPC target to ensure that peak demand for the heating network is given priority response. Critical SOC conditions: When the SOC of any energy storage unit falls below the 20% threshold or rises above the 90% threshold, it is immediately identified as a critical SOC condition. An SOC trend penalty term, set at 10 times the normal weight, is added to the objective function to forcibly bring the SOC back to a safe range. If the BESS reaches its limit, the peak load is automatically switched from the BESS to the pumped storage unit, triggering a local alarm and compensation strategy. The SOC limit violation term is incorporated into the ADMM iterative constraint as a slack variable. Communication delay fluctuation condition: When the estimated communication delay within the sliding two-minute window fluctuates between 100ms and 500ms and the variance exceeds 50ms 2When the communication delay fluctuates, it is identified as a working condition. The sliding window least squares method is used to update the network delay estimate online every 30 seconds, and the current delay is introduced into the MPC prediction equation as a first-order lag mapping term. If the delay exceeds 300ms, the consensus weight is immediately reduced to 0.5 and the number of iterations is reduced to 3 to prioritize local optimization execution.

[0008] The above operations limit the fine-grained configuration of prediction window length, optimized weight distribution, dynamic delay compensation and safety constraints for five typical operating conditions: smooth operation, rapid fluctuations, peak load of the heating network, SOC criticality and communication delay fluctuations. Combined with the dynamic mapping of heat pumps and online prediction of meteorological and historical heat usage data, it further strengthens the support for the accuracy of thermal control.

[0009] To ensure rapid tracking and stable compensation of dynamic changes in network delay: As a further solution of the present invention, in step B, the online delay identification module in the communication delay fluctuation condition adopts the recursive least squares method with a forgetting factor of 0.95 to perform weighted estimation of the communication delay samples in the last 120 seconds, and maps the identified dynamic delay parameters to the delay compensation coefficient in the MPC prediction equation through the first-order inertia model to accurately compensate for network jitter and delay fluctuations.

[0010] To ensure precise matching of power output and heat network demand within the same optimization framework, thereby maximizing the overall energy efficiency of the system, a high-weight penalty is imposed on the state of charge of energy storage units to prevent safety hazards and life degradation caused by overcharging or over-discharging, ensuring that the system maintains safe and reliable operation during the dynamic peak and frequency regulation and heat supply coupling process: As a further solution of the present invention, in step C, the objective function additionally adds an electric-thermal coupling term that multiplies the power output of the battery energy storage unit and the pumped storage unit by their corresponding electric-thermal conversion efficiency, and uses the square of the difference between the deviation from the predicted load of the heat network as a penalty to achieve the coordinated optimization of power output and heat demand. The objective function also additionally adds an SOC constraint term. When the SOC of any energy storage unit is lower than 20% or higher than 90%, the penalty term is calculated as the square of the absolute value of the SOC deviation from the safe area, and the penalty weight is not less than 10.

[0011] To prioritize rapid suppression of frequency disturbances and improve system dynamic response speed and stability when renewable energy output or load fluctuates sharply, while maintaining a long forecast horizon under stable operating conditions to balance dispatch efficiency and economy: As a further solution of the present invention, in step D, in each calculation cycle, the controller first detects the fluctuation amplitude of the renewable energy output and load in the last minute, and compares it with a preset fluctuation threshold. If the preset threshold is exceeded, the controller proportionally shortens the time range used for model prediction based on the current fluctuation size, raises the priority of frequency deviation in the optimization objectives to the highest level, and correspondingly lowers the priority of frequency deviation, energy storage charge state deviation, and heating network temperature deviation. If it is below the threshold, the controller automatically restores the prediction time to the original set long period and restores the priority of each objective to the initial setting.

[0012] To track and compensate for control command delays caused by network latency jitter and packet loss in real time, and ensure that model predictive control instructions can be synchronously issued and executed in a distributed network environment, the timeliness, stability, and robustness of system peak and frequency regulation and heat supply scheduling are improved: As a further solution of the present invention, in step E, the delay identification module uses the weighted sliding window least squares method to perform real-time estimation of the communication delay and packet loss in the last 120 seconds, and automatically resamples when the packet loss rate exceeds 5%; the obtained delay estimate is smoothed by the exponential weighted average method to obtain a dynamic delay parameter; in the model prediction algorithm, the dynamic delay parameter is used as the time constant of the first-order inertial delay model, and the control input is compensated according to the corresponding delay and then used for optimization solution to offset the instruction delay caused by network jitter and packet loss.

[0013] In a distributed architecture and limited communication environment, the system aims to achieve global and consistent coordination of power generation and heating power among regional controllers, ensuring global power and heat balance while also being robust to packet loss and delay jitter. This ensures the real-time, stable, and optimal performance of peak-shaving, frequency regulation, and heating scheduling. As a further solution of the present invention, in step F, the local decision variables of each region include the active power of the unit, the charging and discharging power of the energy storage unit, and the thermal power regulation of the heating network. The active power increment and thermal power increment to be coordinated between the regions are defined as global consensus variables. Each regional controller introduces a Lagrangian dual term and a penalty factor term into the local objective function. The Lagrangian dual term is composed of the difference between the local decision variable and the global consensus variable and the product of the corresponding dual variable. The penalty factor term is composed of the square of the difference between the local decision variable and the global consensus variable multiplied by the penalty factor, so as to form an augmented Lagrangian function. The initial value of the penalty factor is within a preset range and is dynamically adjusted during the iteration process based on the ratio of the local residual to the dual residual (when the local residual is ten times greater than the dual residual, the penalty factor is doubled; when the dual residual is ten times greater than the local residual, the penalty factor is halved. The iterative process includes: solving the corresponding augmented Lagrangian subproblem in parallel in all regions; summarizing the solutions to each subproblem and averaging them to update the global consensus variable; updating the dual variable; the maximum number of iterations is ten, and the iteration is terminated early when both the local residual and the dual residual are lower than a preset threshold). Each regional controller exchanges local decision variables, global consensus variables, and dual variables via IEC61850 GOOSE messages, with an exchange period of no more than 100ms. The global consensus variables and dual variables are initially set to zero vectors, and in the event of communication packet loss or a single message delay exceeding 200ms, the last valid value is used to continue the iteration.

[0014] In the context of large-scale renewable energy grid integration and the parallel development of centralized heating, power system frequency control and heating network temperature regulation are mutually constrained. A single "calculate first, then execute" model cannot achieve real-time coordination between the two. The two regional controllers communicate in real time via IEC61850 GOOSE. Network jitter, packet loss, and latency fluctuations seriously threaten the collaborative effect. One-time optimization alone cannot guarantee command synchronization and stable convergence. Renewable energy output fluctuates rapidly, the dynamics of energy storage units and heating networks differ greatly, and weather, load, and equipment status are constantly changing. To address this, the following specific improvements have been made: As a further solution of the present invention, steps A to F are mapped and fed back to each other according to the following cyclical logic: Step 1, feature encoding: Map the power system frequency, power distribution, energy storage unit SOC and heating network temperature state variables collected in step A into low-dimensional feature vectors; Step 2: Generate initial control instructions: Based on the low-dimensional feature vectors from step 1, generate initial control instructions in step B. The initial control instructions include battery energy storage discharge / charge power, pumped storage unit output, and heat supply distribution of the heating network; Step 3, constraint correction: In step C, the initial control quality is corrected based on the electro-thermal coupling and SOC constraints to obtain a corrected control vector; Step 4, adaptive parameter adjustment: Step D adaptively adjusts the prediction window length and optimization weight according to the output of renewable energy and load fluctuations, and refines the parameters of the corrected control vector to form an adaptive control vector; Step 5, delay compensation: Step E performs dynamic delay identification and compensation on the adaptive control vector and outputs the compensated control vector; Step 6, distributed consensus: Step F uses the ADMM consensus algorithm to solve the compensated control vector in parallel and iteratively to obtain the final control result for each area; Step 7, Response Prediction and Deviation Feedback: The final control result is input into the system model to generate a predicted response. This is then compared with the expected frequency, power, SOC, and temperature targets. The deviation vector between the predicted response and the target is calculated and fed back to the mapping unit as a regularization constraint for the next round of low-dimensional feature vector updates. Step 8, cyclic convergence: repeat steps 1 to 7 until all deviations converge to the preset threshold.

[0015] In this system, due to the rapid fluctuations in wind power and photovoltaic output, measurement noise in the heating network temperature sensor, and model uncertainty caused by the head and pipeline water hammer effect of the pumped storage unit, if only deterministic mapping is used, the generated control instructions will produce severe jitter with the input noise, resulting in frequent switching of energy storage units and increased equipment wear; and in traditional control methods, communication delays and packet loss in multi-region distributed iterations make the information exchanged by each controller incomplete, easily falling into local optimality, slow iteration convergence, and overfitting historical load / heat load curves, resulting in insufficient generalization performance during seasonal or sudden weather changes, making it difficult to maintain dual frequency and temperature steady states.

[0016] As a further solution of the present invention, the mapping unit in step 7 specifically includes an encoder, a sampling unit, a decoder and a distribution regularization module. The encoder inputs the state quantity collected in step A to extract features in the form of a multi-layer neural network, and outputs the mean and variance parameters of the posterior distribution of the latent variable space. The sampling unit generates latent content through random sampling operations based on the mean and variance parameters output by the encoder and independent and identically distributed random noise. The decoder maps the latent vector back to the control instruction space to generate the initial control vector required for step B. The distribution regularization module constrains the posterior distribution parameters by minimizing the divergence between the posterior distribution and the predefined prior distribution.

[0017] The technical effects of the peak-shaving and frequency-modulating heating method for a multi-element energy storage coupled coal-fired power unit of the present invention are as follows: The present invention constructs a dual-time-domain multi-mode electric-thermal hybrid time-domain model that takes into account both millisecond-level frequency response and minute-level energy scheduling under the DC-DMPC distributed model predictive control framework. By combining the SOC-voltage dynamic coupling state equation based on the first-order RC circuit with the nonlinear efficiency curve, temperature-dependent internal resistance and capacity attenuation model, the adaptive optimization of the discharge / charging loss and response speed of battery energy storage and pumped storage is achieved; at the same time, the electric-thermal coupling penalty term and high-weight SOC constraint are introduced into the MPC objective function to ensure efficient matching of power output and heat network load and prevent over-limit of energy storage; the present invention also utilizes real-time online delay identification and first-order inertia compensation, with The encoding-sampling-decoding-regularization cyclic mapping with adaptive prediction window and weight adjustment, as well as the global consensus iteration mechanism based on ADMM, accurately compensate for network delay jitter and dynamically coordinate the heating strategies of batteries, energy storage units and heat pumps in response to typical fluctuations in wind power and photovoltaic power, sudden changes in thermal load and communication jitter and packet loss environments; through multiple rounds of deviation feedback and loop optimization, the coordinated convergence of frequency deviation, power balance, SOC and heating network temperature deviation is achieved, greatly improving the system's dynamic robustness, peak-shaving and frequency regulation and heating coordination efficiency, supporting real-time online adaptive model updates, while significantly reducing system operation risks and extending the life of energy storage and heat pump equipment, achieving safe and efficient long-term continuous operation. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 A diagram of the human-machine control interface for applying the method proposed in the present invention. DETAILED DESCRIPTION

[0019] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0020] Example 1. Figure 1 As shown, the present invention proposes a peak-shaving and frequency-regulating heating method for a multi-element energy storage coupled coal-fired power unit, comprising the following steps: A. Build a two-region interconnected power-thermal system model that includes battery energy storage units, pumped storage units, and heating network loads; B. Design a distributed model predictive controller based on DC-DMPC, with frequency deviation, power balance, energy storage unit SOC, and heating network temperature deviation as optimization objectives; C. Introducing the electrical-thermal coupling term and the energy storage unit SOC constraint into the MPC objective function; D. The controller adaptively adjusts the prediction window length and weighting coefficient based on renewable energy output and load fluctuations; E. Adopt a dynamic delay identification strategy for network jitter and packet loss, and embed a delay compensation model into the prediction model; F. Each regional controller coordinates the power and heat between units through the ADMM consensus algorithm and uses the alternating direction multiplier method to decompose and solve.

[0021] exist Figure 1 In the platform, for the peak-shaving and frequency-regulating heating method for multi-energy storage coupled coal-fired power units proposed in this invention, the platform dynamically displays real-time data such as active power output, heating power of the heat network, system frequency and SOC of the energy storage unit in the "System Status Overview" area, and superimposes the actual power and predicted power lines in the "Real-time Power Curve" to verify the accuracy of the model online; in the "System Operating Condition" panel, the communication delay identification parameters (forgetting factor, delay estimation, packet loss rate), sampling period and frequency and power stability indicators are presented in real time, and the current operating conditions are indicated at the same time; in the "Control Strategy and Parameters" area, the prediction window length and optimization target weights (frequency deviation, power deviation, SOC deviation, temperature deviation), ADMM consensus algorithm relaxation factor are visually adjusted in the form of sliders. The system displays the number of sub-items and iterations, and displays the status of the first-order time delay compensation model; the "Heating Network Temperature Response" and "Multi-Region Power Distribution" charts reflect in real time the heating network regional temperature deviation response rate, pipeline flow rate, and the power distribution ratio of battery energy storage and pumped storage units in each region; and the "Real-time Alarm" and "Operation Log" areas centrally display alarm information and adjustment records such as excessive energy storage temperature, frequency limit violation, abnormal heating network temperature, and abnormal communication. This realizes a closed-loop iteration from prediction parameter setting to algorithm parameter optimization, and then to execution effect monitoring and alarm handling. The system can achieve multi-indicator coordinated adaptive online adjustment and visual management and control based on multiple working conditions such as wind power and photovoltaic fluctuations, sudden changes in thermal load, and network jitter, ensuring efficient, stable and safe peak-shaving and frequency regulation and heating scheduling.

[0022] In order to significantly shorten the frequency response time of the system under the disturbance of renewable energy, improve the discharge / charge efficiency of the energy storage unit, accurately reproduce the spatiotemporal dynamic process of the hydro-electromechanical and thermal network, and ensure that the system frequency stability and the temperature deviation of the thermal network are effectively controlled under the conditions of large-scale grid connection and sudden heat load: In step A, based on the classic first-order RC circuit SOC-voltage model, combined with the nonlinear efficiency curve and temperature-dependent internal resistance calibrated on site, the temperature, SOC and power coupling are dynamically introduced into the state equation, and the discharge / charge loss and response speed under different working conditions are adaptively optimized. The capacity attenuation model based on the cumulative number of cycles and depth is incorporated, and the attenuation rate is set as the decision variable; in order to take into account the millisecond-level frequency regulation and minute-level energy scheduling requirements, a hybrid is constructed using dual discrete steps of 0.1s and 60s. A time-domain model is developed, and a dual-mode state equation is designed for both pumping and power generation modes. A hydraulic-electromechanical coupling term containing head changes and a first-order lag pipeline water hammer approximation term are embedded to capture pressure oscillations and power jitters when flow rates suddenly change. At the same time, the effective head and water output capacity are dynamically calculated based on the height difference between the two areas and the pipeline length, and coordinated with the grid demand. In terms of the heating network, the regional heating network is divided into multi-node pipe sections to establish heat capacity and thermal resistance equations, simulate the spatiotemporal diffusion of pipeline heat loss, flow velocity and supply and return water temperature difference, and allow MPC to directly adjust the power output of the battery energy storage unit and the pumped storage unit in the dynamic mapping model of the heat pump unit to meet the temperature and heat demand of the heating network. Combined with local meteorological and historical heat consumption curves, the recursive least squares method is used to update the load forecasting model online and quickly correct sudden heating demands.

[0023] A two-region interconnected power-thermal coupling scenario where wind power, photovoltaic bases, and regional thermal load centers are decentralized and connected through a dual energy storage system of battery energy storage and pumped hydropower storage. This scenario requires simultaneous assurance of system frequency stability, energy storage safety, and thermal network temperature balance during severe cold weather or sharp load fluctuations, while also facing challenges such as network jitter and packet loss.

[0024] It should be noted that step B includes but is not limited to the following specific working conditions: Stable operating condition: A stable operating condition is identified only if the system frequency standard deviation is less than 0.02 Hz, the power deviation standard deviation is less than 5% of the rated power, and the communication delay jitter is consistently less than 50 ms for five consecutive minutes. In this case, the prediction window is fixed at 10 minutes. The controller optimizes the frequency, power, energy storage unit SOC, and heating network temperature deviation using established empirical ratios. The ADMM consensus iteration uses a "lightweight" mode with a relaxation factor of 1 and five iterations. Rapid Fluctuation Conditions: If the sudden change rate of load or renewable energy output exceeds 20% of the rated power within any minute, or the absolute value of the frequency slope is greater than 0.05 Hz / s, it is determined to be a rapid fluctuation condition. In this case, the prediction window is shortened to 5 minutes, the frequency deviation weight is increased to 2, and the remaining weights are adjusted proportionally. The ADMM consensus algorithm automatically increases the relaxation factor from 1 to 5 and the number of iterations to 8 based on the local residual. If the local MPC solution exceeds 0.2 seconds, the last available solution is automatically used to ensure real-time response. Heating network peak load condition: When the return water temperature deviation exceeds 3°C for three consecutive minutes and the heat load growth rate exceeds 10% of the rated heat within five minutes, it is identified as a heating network peak load condition. At this time, the heating network temperature deviation weight is increased to 3, and the frequency deviation weight is reduced to 0.5. A dual-channel recursive least squares forecast based on weather forecasts and historical heat consumption curves is run in parallel, and a heat lower limit constraint is added to the MPC target to ensure that peak demand for the heating network is given priority response. Critical SOC conditions: When the SOC of any energy storage unit falls below the 20% threshold or rises above the 90% threshold, it is immediately identified as a critical SOC condition. An SOC trend penalty term, set at 10 times the normal weight, is added to the objective function to forcibly bring the SOC back to a safe range. If the BESS reaches its limit, the peak load is automatically switched from the BESS to the pumped storage unit, triggering a local alarm and compensation strategy. The SOC limit violation term is incorporated into the ADMM iterative constraint as a slack variable. Communication delay fluctuation condition: When the estimated communication delay within the sliding two-minute window fluctuates between 100ms and 500ms and the variance exceeds 50ms 2 When the communication delay fluctuates, it is identified as a working condition. The sliding window least squares method is used to update the network delay estimate online every 30 seconds, and the current delay is introduced into the MPC prediction equation as a first-order lag mapping term. If the delay exceeds 300ms, the consensus weight is immediately reduced to 0.5 and the number of iterations is reduced to 3 to prioritize local optimization execution.

[0025] For example, the system platform monitored the following data from 08:30 to 08:35 on a certain day: frequency 50.01 Hz ± 0.004 Hz (5-minute standard deviation 0.004 Hz < 0.02 Hz), power deviation ± 2.0 MW (standard deviation 2.5% < 5%), average communication jitter 30 ms, and jitter amplitude < 50 ms. These data meet the following criteria: "frequency standard deviation < 0.02 Hz, power deviation < 5%, and average communication jitter < 50 ms for 5 consecutive minutes." This indicates a stable operating condition. The application strategy is a 10-minute prediction window with weights (frequency: power: SOC: Temperature = 1:1:0.5:0.5). The ADMM consensus iteration uses a relaxation factor of 1 and is repeated five times. The controller performs only the minimum number of iterations, freeing up communication bandwidth and optimizing all indicators in a balanced manner, maintaining stability at the lowest cost. Within the window of 08:36-08:37, the frequency suddenly changes from 50Hz to 48.5Hz, with a mutation rate of 1.5Hz / s to 0.05Hz / s. The window is shortened to 5 minutes, the frequency weight is increased to 2, and other weights are simultaneously decreased. The ADMM consensus iteration relaxation factor is dynamically increased to 5, and eight iterations are performed. The MPC solution time is 0.2s, and the last solution is retained for immediate execution. Similarly, specific instances of heating network peak load conditions, critical SOC conditions, and fluctuating communication delay conditions are identified through identification conditions and executed according to the aforementioned adaptive application strategies.

[0026] The above operations limit the fine-grained configuration of prediction window length, optimized weight distribution, dynamic delay compensation and safety constraints for five typical operating conditions: smooth operation, rapid fluctuations, peak load of the heating network, SOC criticality and communication delay fluctuations. Combined with the dynamic mapping of heat pumps and online prediction of meteorological and historical heat usage data, it further strengthens the support for the accuracy of thermal control.

[0027] To ensure rapid tracking and stable compensation for dynamic changes in network delay, in step B, the online delay identification module in the communication delay fluctuation condition uses a recursive least squares method with a forgetting factor of 0.95 to perform a weighted estimate of the communication delay samples within the last 120 seconds. The identified dynamic delay parameters are mapped into delay compensation coefficients in the MPC prediction equation through a first-order inertia model to accurately compensate for network jitter and delay fluctuations. The recursive least squares (RLS) method with a forgetting factor of 0.95 can give higher weight to the latest delay samples, respond quickly to the most recent network conditions, and gradually attenuate the impact of past data. When the communication delay changes suddenly due to temporary jitter or network packet jitter, the system can accurately capture the new delay level within tens of seconds, providing a reliable basis for subsequent compensation; the delay parameters obtained by online identification are then smoothed through the first-order inertia model, which is equivalent to a filtering process between high-frequency noise and low-frequency trends. This ensures that the compensation parameters will not oscillate significantly due to measurement noise or short-term jitter, and also retains sensitivity to real network delay mutations, thereby avoiding over-compensation or failure to keep up with changes. The dilemma of digitization; introducing a real-time updated delay compensation coefficient into the prediction equation of model predictive control (MPC) can effectively eliminate the misalignment between "control command and actual execution" caused by communication delay, so that the control input can act on the on-site device in a timely and accurate manner, greatly reducing the frequency deviation and heating network temperature error caused by delay; network jitter and packet loss often interrupt the pace consistency between distributed controllers, resulting in convergence difficulties or local oscillations. Through dynamic delay identification and inertia compensation, distributed MPC can maintain high synchronization accuracy under different delay conditions, reduce dependence on conservative parameter settings, thereby accelerating algorithm convergence and improving the robustness of overall peak-shaving, frequency regulation and heating control.

[0028] To ensure that power output and heat network demand can be precisely matched within the same optimization framework, thereby maximizing the overall energy utilization efficiency of the system, a high-weight penalty is imposed on the state of charge of the energy storage unit to prevent safety hazards and life degradation caused by overcharging or over-discharging, ensuring that the system always maintains a safe and reliable operating state during the dynamic peak-frequency regulation and heat supply coupling process. In step C, the additional electric-thermal coupling term added to the objective function multiplies the power output of the battery energy storage unit and the pumped storage unit by the corresponding electric-thermal conversion efficiency, and uses the square of the difference between the deviation from the predicted load of the heat network as a penalty to achieve the coordinated optimization of power output and heat demand. The additional SOC constraint term added to the objective function calculates the penalty term as the square of the absolute value of the SOC deviation from the safe area when the SOC of any energy storage unit is lower than 20% or higher than 90%, and the penalty weight is not less than 10.

[0029] In order to prioritize the rapid suppression of frequency disturbances and improve the system's dynamic response speed and stability when renewable energy output or load fluctuates violently, while maintaining a longer prediction horizon under stable operating conditions to balance scheduling efficiency and economy, in step D, in each operation cycle, the controller first detects the fluctuation amplitude of renewable energy output and load in the last minute and compares it with the pre-set fluctuation threshold. If it exceeds the pre-set threshold, the controller proportionally shortens the time range used for model prediction based on the current fluctuation size, raises the priority of frequency deviation in the optimization objective to the highest level, and correspondingly lowers the priority of frequency deviation, energy storage charge state deviation, and heating network temperature deviation. If it is below the threshold, the controller automatically restores the prediction time to the original set long period and restores the priority of each objective to the initial setting.

[0030] To track and compensate for control command delays caused by network jitter and packet loss in real time, and to ensure that model predictive control instructions can be synchronously issued and executed in a distributed network environment, thereby improving the timeliness, stability, and robustness of the system's peak-shaving, frequency regulation, and heat supply scheduling, in step E, the delay identification module uses the weighted sliding window least squares method to perform real-time estimation of the communication delay and packet loss for the last 120 seconds, and automatically resamples when the packet loss rate exceeds 5%. The resulting delay estimate is smoothed using the exponentially weighted average method to obtain a dynamic delay parameter. In the model prediction algorithm, this dynamic delay parameter is used as the time constant of the first-order inertial delay model. The control input is compensated according to the corresponding delay and then used in the optimization solution to offset the command delay caused by network jitter and packet loss.

[0031] In order to achieve global and consistent coordination of power generation and thermal power among regional controllers in a distributed architecture and restricted communication environment, ensuring global power and heat balance, and being robust to packet loss and delay jitter, thereby ensuring the real-time, stability and optimality of peak-shaving, frequency regulation and heating scheduling, in step F, the local decision-making quantities of each region include the active power of the unit, the charging and discharging power of the energy storage unit and the thermal power regulation quantity of the heating network. The active power increment and thermal power increment to be coordinated between the regions are defined as global consensus variables; each regional controller introduces the Lagrangian dual term and the penalty factor term in the local objective function. The Lagrangian dual term is composed of the difference between the local decision-making quantity and the global consensus variable and the product of the corresponding dual variable. The penalty factor term is the product of the difference between the local decision-making quantity and the global consensus variable and the corresponding dual variable. The penalty factor term is composed of the square of the difference between the local decision variable and the global consensus variable multiplied by the penalty factor to form an augmented Lagrangian function. The initial value of the penalty factor is within a preset range and is dynamically adjusted during the iteration process according to the ratio of the local residual to the dual residual (when the local residual is ten times greater than the dual residual, the penalty factor is doubled; when the dual residual is ten times greater than the local residual, the penalty factor is halved. The iterative process includes: solving the corresponding augmented Lagrangian subproblem in parallel in all regions; summarizing the solutions of each subproblem and averaging them to update the global consensus variable; updating the dual variable; the maximum number of iterations is ten, and the system terminates early when both the local residual and the dual residual are lower than a preset threshold). Each regional controller exchanges local decision variables, global consensus variables, and dual variables via IEC61850 GOOSE messages, with an exchange period of no more than 100ms. The global consensus variable and the dual variable are initially set to zero vectors. In the event of communication packet loss or a single message delay exceeding 200ms, the iteration continues using the last valid value.

[0032] The present invention achieves precise matching of power output and heat network load by introducing electric-thermal coupling penalty and high-weight SOC constraint in the same optimization framework, and ensures rapid response of the system under severe fluctuations and efficient operation under stable working conditions through adaptive prediction window and weight adjustment; at the same time, online delay identification and inertia compensation eliminate the impact of network jitter and packet loss on control synchronization, and the ADMM consensus iteration mechanism ensures global consistency of power and heat in each region; unlike the existing technology that relies on fixed time domain and static weights and cannot take into account both electric-thermal coordination and delay jitter compensation, the present invention takes into account dynamic performance, operational safety and communication robustness, significantly improving the system's energy utilization efficiency, peak-shaving and frequency regulation and heat supply coordination capabilities, and long-term stable operation reliability.

[0033] Example 2. The difference between Example 2 of the present invention and Example 1 is that this example introduces a peak-shaving and frequency-regulating heating method for a multi-element energy storage coupled coal-fired power unit.

[0034] In the context of large-scale renewable energy grid integration and the parallel development of centralized heating, power system frequency control and heating network temperature regulation are mutually constrained. A single "calculate first, then execute" model cannot achieve real-time coordination between the two. The two regional controllers communicate in real time via IEC61850 GOOSE. Network jitter, packet loss, and latency fluctuations seriously threaten the collaborative effect. One-time optimization alone cannot guarantee command synchronization and stable convergence. Renewable energy output fluctuates rapidly, the dynamics of energy storage units and heating networks differ greatly, and weather, load, and equipment status are constantly changing. To address this, the following specific improvements have been made: First, steps A to F are mapped and fed back to each other according to the following cyclical logic: Step 1, feature encoding: Map the power system frequency, power distribution, energy storage unit SOC and heating network temperature state variables collected in step A into low-dimensional feature vectors; Step 2: Generate initial control instructions: Based on the low-dimensional feature vectors from step 1, generate initial control instructions in step B. The initial control instructions include battery energy storage discharge / charge power, pumped storage unit output, and heat supply distribution of the heating network; Step 3, constraint correction: In step C, the initial control quality is corrected based on the electro-thermal coupling and SOC constraints to obtain a corrected control vector; Step 4, adaptive parameter adjustment: Step D adaptively adjusts the prediction window length and optimization weight according to the output of renewable energy and load fluctuations, and refines the parameters of the corrected control vector to form an adaptive control vector; Step 5, delay compensation: Step E performs dynamic delay identification and compensation on the adaptive control vector and outputs the compensated control vector; Step 6, distributed consensus: Step F uses the ADMM consensus algorithm to solve the compensated control vector in parallel and iteratively to obtain the final control result for each area; Step 7, Response Prediction and Deviation Feedback: The final control result is input into the system model to generate a predicted response. This is then compared with the expected frequency, power, SOC, and temperature targets. The deviation vector between the predicted response and the target is calculated and fed back to the mapping unit as a regularization constraint for the next round of low-dimensional feature vector updates. Step 8, cyclic convergence: repeat steps 1 to 7 until all deviations converge to the preset threshold.

[0035] In this system, due to the rapid fluctuations in wind power and photovoltaic output, measurement noise in the heating network temperature sensor, and model uncertainty caused by the head and pipeline water hammer effect of the pumped storage unit, if only deterministic mapping is used, the generated control instructions will produce severe jitter with the input noise, resulting in frequent switching of energy storage units and increased equipment wear; and in traditional control methods, communication delays and packet loss in multi-region distributed iterations make the information exchanged by each controller incomplete, easily falling into local optimality, slow iteration convergence, and overfitting historical load / heat load curves, resulting in insufficient generalization performance during seasonal or sudden weather changes, making it difficult to maintain dual frequency and temperature steady states.

[0036] Therefore, in order to further improve the above mechanism, the mapping unit in step 7 specifically includes an encoder, a sampling unit, a decoder and a distribution regularization module. The encoder inputs the state quantity collected in step A to extract features in the form of a multi-layer neural network, and outputs the mean and variance parameters of the posterior distribution of the latent variable space. The sampling unit generates latent content through random sampling operations based on the mean and variance parameters output by the encoder and independent and identically distributed random noise. The decoder maps the latent vector back to the control instruction space to generate the initial control vector required for step B. The distribution regularization module constrains the posterior distribution parameters by minimizing the divergence between the posterior distribution and the predefined prior distribution.

[0037] By introducing an end-to-end joint optimization mechanism of encoder-sampling unit-decoder-distribution regularization module in the mapping unit, the high-dimensional grid frequency, power distribution, energy storage SOC and heat network temperature state variables are mapped into parameterized latent distributions. Diverse control vectors are generated by random sampling in the implicit space and the posterior and prior distribution divergences are minimized through distribution regularization, thereby suppressing the high-frequency jitter caused by measurement noise and model uncertainty, and avoiding the control instructions caused by uncertainty mapping from drastic fluctuations with input disturbances. At the same time, the predicted response and target difference are regularized and mapped back to the latent space with the help of cyclic deviation feedback, realizing online adaptive model update, which significantly improves The synchronization and convergence speed of distributed MPC in environments with communication jitter, packet loss and delay effectively prevent traditional multi-region optimization methods from falling into local optimality or slow convergence due to incomplete information. Through multiple rounds of mapping-correction-decoding iterations, it maintains multi-objective collaborative convergence of frequency regulation, power balance, energy storage safety and heating network temperature under multiple working conditions such as rapid fluctuations in wind power and photovoltaics, sudden loads in the heating network and dynamic differences in equipment, and ultimately achieves a significant improvement in the system's dynamic robustness, energy utilization efficiency and thermal accuracy, providing reliable, efficient and sustainable technical support for the electric-thermal coupled peak-shaving and frequency-regulating heating method in the scenario of large-scale renewable energy grid connection and centralized heating in parallel.

[0038] Combining Example 1 and Example 2, the present invention constructs a dual-time-domain multi-mode electric-thermal hybrid time-domain model that takes into account both millisecond-level frequency response and minute-level energy scheduling under the DC-DMPC distributed model predictive control framework. By combining the SOC-voltage dynamic coupling state equation based on the first-order RC circuit with the nonlinear efficiency curve, temperature-dependent internal resistance and capacity attenuation model, adaptive optimization of battery energy storage and pumped storage discharge / charging loss and response speed is achieved; at the same time, electric-thermal coupling penalty terms and high-weight SOC constraints are introduced into the MPC objective function to ensure efficient matching of power output and heat network load and prevent over-limit of energy storage; the present invention also utilizes real-time online delay identification and first-order Inertia compensation, encoding-sampling-decoding-regularization cyclic mapping with adaptive prediction window and weight adjustment, and global consensus iteration mechanism based on ADMM, accurately compensate for network delay jitter and dynamically coordinate battery, energy storage unit and heat pump heating strategies for typical fluctuations in wind power and photovoltaic, sudden changes in thermal load and communication jitter and packet loss environments; through multiple rounds of deviation feedback and loop optimization, the coordinated convergence of frequency deviation, power balance, SOC and heating network temperature deviation is achieved, greatly improving the system's dynamic robustness, peak-shaving and frequency regulation and heating coordination efficiency, supporting real-time online adaptive model updates, while significantly reducing system operation risks and extending the life of energy storage and heat pump equipment, achieving safe and efficient long-term continuous operation.

[0039] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.

[0040] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A peak-shaving and frequency-modulating heating method for a multi-element energy storage coupled coal-fired power unit, characterized in that: The steps include: A. Build a two-region interconnected power-thermal system model that includes battery energy storage units, pumped storage units, and heating network loads; B. Design a distributed model predictive controller based on DC-DMPC, with frequency deviation, power balance, energy storage unit SOC, and heating network temperature deviation as optimization objectives; C. Introducing the electrical-thermal coupling term and the energy storage unit SOC constraint into the MPC objective function; D. The controller adaptively adjusts the prediction window length and weighting coefficient based on renewable energy output and load fluctuations; E. Adopt a dynamic delay identification strategy for network jitter and packet loss, and embed a delay compensation model into the prediction model; F. Each regional controller coordinates the power and heat between units through the ADMM consensus algorithm and uses the alternating direction multiplier method to decompose and solve.

2. A peak-shaving and frequency-modulating heating method for a multi-element energy storage coupled coal-fired power unit according to claim 1, characterized in that: In step A, based on the classic first-order RC circuit SOC-voltage model, combined with the nonlinear efficiency curve and temperature-dependent internal resistance calibrated on site, the temperature, SOC and power coupling are dynamically introduced into the state equation, the discharge / charge loss and response speed under different working conditions are adaptively optimized, and the capacity decay model based on the cumulative number of cycles and depth is incorporated, and the decay rate is set as the decision variable; in order to take into account the millisecond-level frequency regulation and minute-level energy scheduling requirements, a hybrid time domain model is constructed with dual discrete step lengths of 0.1s and 60s, and a dual-mode state equation is designed for both pumping and power generation modes, with an embedded hydraulic-electromechanical coupling term containing water head changes and a first-order lag tube. The water hammer approximation term is used to capture the pressure oscillation and power jitter when the flow rate changes suddenly. At the same time, the effective head and water output capacity are dynamically calculated according to the height difference between the two areas and the pipeline length, and coordinated with the grid demand. In terms of the heating network, the regional heating network is divided into multi-node pipe sections to establish heat capacity and thermal resistance equations, simulate the spatiotemporal diffusion of pipeline heat loss, flow rate and supply and return water temperature difference, and allow MPC to directly adjust the power output of battery energy storage units and pumped storage units in the dynamic mapping model of heat pump units to meet the temperature and heat demand of the heating network. Combined with local meteorological and historical heat consumption curves, the recursive least squares method is used to update the load forecasting model online and quickly correct sudden heating demand.

3. The peak-shaving and frequency-modulating heating method for a multi-element energy storage coupled coal-fired power unit according to claim 1, characterized in that: Step B includes but is not limited to the following specific working conditions: Stable operating condition: A stable operating condition is identified only if, for five consecutive minutes, the system frequency standard deviation is less than 0.02Hz, the power deviation standard deviation is less than 5% of the rated power, and the communication delay jitter is consistently less than 50ms. In this case, the prediction window is fixed at 10 minutes. The controller optimizes the frequency, power, energy storage unit SOC, and heating network temperature deviation using established empirical ratios. The ADMM consensus iteration uses a "lightweight" mode with a relaxation factor of 1 and five iterations. Rapid Fluctuation Conditions: If the sudden change rate of load or renewable energy output exceeds 20% of the rated power within any minute, or the absolute value of the frequency slope is greater than 0.05 Hz / s, it is determined to be a rapid fluctuation condition. In this case, the prediction window is shortened to 5 minutes, the frequency deviation weight is increased to 2, and the remaining weights are adjusted proportionally. The ADMM consensus algorithm automatically increases the relaxation factor from 1 to 5 and the number of iterations to 8 based on the local residual. If the local MPC solution exceeds 0.2 seconds, the last available solution is automatically used to ensure real-time response. Heating network peak load condition: When the return water temperature deviation exceeds 3°C for three consecutive minutes and the heat load growth rate exceeds 10% of the rated heat within five minutes, it is identified as a heating network peak load condition. At this time, the heating network temperature deviation weight is increased to 3, and the frequency deviation weight is reduced to 0.

5. A dual-channel recursive least squares forecast based on weather forecasts and historical heat consumption curves is run in parallel, and a heat lower limit constraint is added to the MPC target to ensure that peak demand for the heating network is given priority response. Critical SOC conditions: When the SOC of any energy storage unit falls below the 20% threshold or rises above the 90% threshold, it is immediately identified as a critical SOC condition. An SOC trend penalty term, set at 10 times the normal weight, is added to the objective function to forcibly bring the SOC back to a safe range. If the BESS reaches its limit, the peak load is automatically switched from the BESS to the pumped storage unit, triggering a local alarm and compensation strategy. The SOC limit violation term is incorporated into the ADMM iterative constraint as a slack variable. Communication delay fluctuation condition: When the estimated communication delay within the sliding two-minute window fluctuates between 100ms and 500ms and the variance exceeds 50ms 2 When the communication delay fluctuates, it is identified as a working condition. The sliding window least squares method is used to update the network delay estimate online every 30 seconds, and the current delay is introduced into the MPC prediction equation as a first-order lag mapping term. If the delay exceeds 300ms, the consensus weight is immediately reduced to 0.5 and the number of iterations is reduced to 3 to prioritize local optimization execution.

4. A peak-shaving and frequency-modulating heating method for a multi-element energy storage coupled coal-fired power unit according to claim 3, characterized in that: In step B, the online delay identification module in the communication delay fluctuation condition uses the recursive least squares method with a forgetting factor of 0.95 to perform weighted estimation of the communication delay samples in the last 120 seconds, and maps the identified dynamic delay parameters into the delay compensation coefficients in the MPC prediction equation through a first-order inertia model to accurately compensate for network jitter and delay fluctuations.

5. The peak-shaving and frequency-modulating heating method for a multi-element energy storage coupled coal-fired power unit according to claim 1, characterized in that: In step C, the objective function adds an additional electro-thermal coupling term that multiplies the power output of the battery energy storage unit and the pumped storage unit by their corresponding electro-thermal conversion efficiency, and uses the square of the difference between the deviation from the predicted load of the heat network as a penalty to achieve the coordinated optimization of power output and heat demand. The objective function also adds an SOC constraint term that calculates the penalty term as the square of the absolute value of the SOC deviation from the safe area when the SOC of any energy storage unit is lower than 20% or higher than 90%, and the penalty weight is no less than 10.

6. The peak-shaving and frequency-modulating heating method for a multi-element energy storage coupled coal-fired power unit according to claim 1, characterized in that: In step D, during each calculation cycle, the controller first detects the fluctuation amplitude of renewable energy output and load in the last minute and compares it with the pre-set fluctuation threshold. If the pre-set threshold is exceeded, the controller proportionally shortens the time range used for model prediction based on the current fluctuation size, raises the priority of frequency deviation in the optimization target to the highest level, and correspondingly lowers the priority of frequency deviation, energy storage charge state deviation, and heating network temperature deviation. If it is below the threshold, the controller automatically restores the prediction time to the original set long period and restores the priority of each target to the initial setting.

7. The peak-shaving and frequency-modulating heating method for a multi-element energy storage coupled coal-fired power unit according to claim 1, characterized in that: In step E, the delay identification module uses the weighted sliding window least squares method to perform real-time estimation of the communication delay and packet loss in the last 120 seconds, and automatically resamples when the packet loss rate exceeds 5%; The obtained delay estimate is smoothed using the exponentially weighted average method to obtain the dynamic delay parameter. In the model prediction algorithm, this dynamic delay parameter is used as the time constant of the first-order inertial delay model. The control input is compensated according to the corresponding delay and then used in the optimization solution to offset the command delay caused by network jitter and packet loss.

8. The peak-shaving and frequency-modulating heating method for a multi-element energy storage coupled coal-fired power unit according to claim 1, characterized in that: In step F, the local decision quantities of each region include the active power of the unit, the charging and discharging power of the energy storage unit, and the thermal power regulation quantity of the heating network. The active power increment and thermal power increment to be coordinated between the regions are defined as global consensus variables; each regional controller introduces a Lagrangian dual term and a penalty factor term into the local objective function. The Lagrangian dual term is composed of the product of the difference between the local decision quantity and the global consensus variable and the corresponding dual variable. The penalty factor term is composed of the square of the difference between the local decision quantity and the global consensus variable multiplied by the penalty factor to form an augmented Lagrangian function. The initial value of the penalty factor is within a preset range and is dynamically adjusted according to the ratio of the local residual to the dual residual during the iteration process.

9. The peak-shaving and frequency-modulating heating method for a multi-element energy storage coupled coal-fired power unit according to claim 1, characterized in that: Steps A to F are mapped and fed back to each other according to the following cyclical logic: Step 1, feature encoding: Map the power system frequency, power distribution, energy storage unit SOC and heating network temperature state variables collected in step A into low-dimensional feature vectors; Step 2: Generate initial control instructions: Based on the low-dimensional feature vectors from step 1, generate initial control instructions in step B. The initial control instructions include battery energy storage discharge / charge power, pumped storage unit output, and heat supply distribution of the heating network; Step 3, constraint correction: In step C, the initial control quality is corrected based on the electro-thermal coupling and SOC constraints to obtain a corrected control vector; Step 4, adaptive parameter adjustment: Step D adaptively adjusts the prediction window length and optimization weight according to the output of renewable energy and load fluctuations, and refines the parameters of the corrected control vector to form an adaptive control vector; Step 5, delay compensation: Step E performs dynamic delay identification and compensation on the adaptive control vector and outputs the compensated control vector; Step 6, distributed consensus: Step F uses the ADMM consensus algorithm to solve the compensated control vector in parallel and iteratively to obtain the final control result for each area; Step 7, Response Prediction and Deviation Feedback: The final control result is input into the system model to generate a predicted response. This is then compared with the expected frequency, power, SOC, and temperature targets. The deviation vector between the predicted response and the target is calculated and fed back to the mapping unit as a regularization constraint for the next round of low-dimensional feature vector updates. Step 8, cyclic convergence: repeat steps 1 to 7 until all deviations converge to the preset threshold.

10. A peak-shaving and frequency-modulating heating method for a multi-element energy storage coupled coal-fired power unit according to claim 9, characterized in that: The mapping unit in step 7 specifically includes an encoder, a sampling unit, a decoder and a distribution regularization module. The encoder inputs the state quantity collected in step A to extract features in the form of a multi-layer neural network, and outputs the mean and variance parameters of the posterior distribution of the latent variable space. The sampling unit generates latent content through random sampling operations based on the mean and variance parameters output by the encoder and independent and identically distributed random noise. The decoder maps the latent vector back to the control instruction space to generate the initial control vector required for step B. The distribution regularization module constrains the posterior distribution parameters by minimizing the divergence between the posterior distribution and the predefined prior distribution.

Images