Zero-carbon park energy intelligent control method and system based on power prediction

Abstract

The present application relates to a zero-carbon park energy intelligent control method and system based on power prediction, comprising the following steps: obtaining park multi-source data, the multi-source data including historical power load data, weather forecast data, renewable energy power generation data and carbon emission monitoring data; based on the multi-source data, performing multi-time scale power prediction through a pre-constructed hybrid prediction model to generate ultra-short-term power prediction sequence, short-term power prediction sequence and medium and long-term power prediction sequence; according to the ultra-short-term power prediction sequence, the short-term power prediction sequence and the medium and long-term power prediction sequence, a park energy system optimization scheduling model considering carbon flow coupling characteristics is constructed; the optimization scheduling model is solved to generate a day-ahead scheduling plan and a real-time correction plan containing source-grid-load-storage multi-end cooperation; the present application can realize more intelligent zero-carbon park energy intelligent control.

Description

Technical Field

[0001] This invention relates to the field of intelligent control, specifically to a zero-carbon industrial park energy intelligent control method and system based on power forecasting. Background Technology

[0002] Zero-carbon industrial parks have become important carriers for energy structure transformation and low-carbon industrial development. The energy systems of these parks are gradually evolving towards a distributed architecture that coordinates multiple terminals (source, grid, load, and storage), integrating renewable energy generation equipment such as photovoltaics and wind power, energy storage systems, and flexible load units. This places higher demands on the intelligence, precision, and low-carbon level of energy management. However, current technological development in the field of intelligent energy control in industrial parks still has many shortcomings, making it difficult to adapt to the actual operational needs of zero-carbon parks. For example, the power forecasting system is imperfect; existing technologies mostly use single-timescale or single-type forecasting models, failing to combine the advantages of physical mechanisms and data-driven models, and failing to integrate carbon emission constraints into the forecasting process. Ultra-short-term forecasts lack real-time correction mechanisms for meteorological changes and photovoltaic power fluctuations; short-term forecasts struggle to accurately capture the correlation between multiple characteristics such as weather and production plans and load; and medium- and long-term forecasts do not align with the park's carbon quota management requirements, resulting in insufficient forecast accuracy at various time scales and an inability to provide reliable data support for different scheduling scenarios within the park.

[0003] The coupling and multi-objective considerations of energy system optimization scheduling are insufficient. Existing scheduling models mostly aim only at minimizing economic costs, without constructing a coupled control system for power flow and carbon flow, lacking node-level carbon flow rate constraints, and carbon emission control is at a rough level. At the same time, the constraint conditions of the scheduling models do not fully take into account the operating characteristics of multiple terminals such as power generation, grid, load and storage. The solution algorithm is prone to problems such as imbalance between global exploration and local development capabilities and premature convergence. It is difficult to meet the multi-objective optimization needs such as comprehensive cost, carbon emissions and renewable energy consumption rate while satisfying the physical constraints of equipment.

[0004] The park's energy control architecture has obvious defects. It mostly adopts a centralized control mode, with a large computing load on the central controller and a slow control response speed. Furthermore, distributed energy devices and energy storage systems are not equipped with localized autonomous control capabilities. Once communication is interrupted, it will directly lead to loss of control and make it impossible to achieve hierarchical distributed collaborative control of source, grid, load and storage. The system's reliability and robustness are insufficient.

[0005] Without establishing a closed-loop iterative optimization mechanism for models and controls, existing technologies have disconnected power prediction models, optimal scheduling models, and control strategies. Model parameters and constraints are mostly fixed and cannot be dynamically adjusted based on the deviation between actual control response data and predicted values. At the same time, there is a lack of accurate identification of the sources of deviation and targeted optimization methods, resulting in poor adaptability of the intelligent control system. With changes in park weather, load, and equipment operating status, prediction accuracy and control effect continue to decline.

[0006] The lack of visualization and traceability capabilities for carbon flow management in industrial parks makes it difficult for existing technologies to achieve real-time tracking and quantitative accounting of carbon flows. The absence of a carbon flow topology network linked to the physical energy system makes it impossible to accurately calculate the carbon footprint of each energy-consuming unit. When carbon emissions exceed the standard, it is difficult to quickly trace back to the source and locate the key path and main contributing equipment. Carbon emission reduction policies lack accurate data support, making it difficult to implement the transformation from passive monitoring to proactive management. Therefore, a zero-carbon industrial park energy intelligent control method and system based on power forecasting is proposed. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides a zero-carbon industrial park energy intelligent control method based on electricity forecasting, comprising the following steps:

[0008] Acquire multi-source data from the park, including historical power load data, weather forecast data, renewable energy power generation data, and carbon emission monitoring data;

[0009] Based on multi-source data, a pre-built hybrid prediction model is used to perform power forecasting at multiple time scales, generating ultra-short-term power forecasting sequences, short-term power forecasting sequences, and medium- and long-term power forecasting sequences.

[0010] Based on ultra-short-term power forecast sequences, short-term power forecast sequences, and medium- and long-term power forecast sequences, an optimal scheduling model for the park's energy system considering carbon flow coupling characteristics is constructed.

[0011] Solve the optimization scheduling model to generate a day-ahead scheduling plan and a real-time correction plan that includes multi-terminal coordination of source, grid, load and storage;

[0012] Based on the day-ahead scheduling plan and the real-time correction plan, the distributed energy equipment, energy storage system and flexible load in the park are controlled in a hierarchical distributed and coordinated manner through the edge computing gateway, and control response data is collected in real time.

[0013] Based on the deviation between the control response data and the corresponding power forecast values, the parameters of the hybrid forecasting model and the constraints of the optimized scheduling model are dynamically adjusted to form a model-control closed-loop iterative optimization.

[0014] Furthermore, the hybrid prediction model combines a physical mechanism model with a data-driven model, wherein:

[0015] The ultra-short-term power forecast sequence is generated using a similar day algorithm with real-time meteorological correction. Specifically, the measured meteorological data within a preset time period before the current moment is dynamically time-normalized and matched with historical meteorological data. Several historical days with the highest similarity are selected as a set of similar days. The power load data of the set of similar days is weighted and averaged to obtain the basic ultra-short-term forecast value. Based on the real-time output data of the photovoltaic inverter, the difference between the sliding average output value within the previous time window and the real-time output is calculated. This difference is used as the ultra-short-term photovoltaic power fluctuation correction amount and superimposed on the basic ultra-short-term forecast value to generate a preliminary ultra-short-term power forecast sequence.

[0016] The short-term power forecast sequence is generated by introducing a bidirectional long short-term memory network with an attention mechanism. The input features of the bidirectional long short-term memory network with the attention mechanism include weather forecast data, date type, historical load data, and park production plan data.

[0017] The medium- and long-term power forecast series is generated by combining a seasonal autoregressive moving average model with carbon emission constraints. The carbon emission constraints are obtained by allocating the total annual carbon quota of the park to each month according to the historical monthly load ratio, and then introducing the monthly allowable power consumption conversion factor into the model.

[0018] Furthermore, the preliminary ultra-short-term power forecast sequence still needs to be corrected by meteorological change response. Specifically, the meteorological station data is monitored in real time, and when the rate of change of light intensity or temperature within a preset time period is detected to exceed the set threshold, the meteorological change response mechanism is triggered.

[0019] Calculate the change in light intensity over a preset time period prior to the current moment. and average light intensity and temperature change and average temperature T avg ;

[0020] rate of change of light intensity With temperature change rate Multiplying each factor by a preset weighting coefficient α and β and then summing the results yields the photovoltaic power output mutation correction coefficient M. pv , where α+β=1;

[0021] Multiply the preliminary ultra-short-term power forecast sequence by the photovoltaic power output mutation correction factor M. pv The corrected ultra-short-term power forecast sequence is obtained.

[0022] Photovoltaic power output mutation correction coefficient The calculation formula is:

[0023] .

[0024] Furthermore, the optimal scheduling model for the park's energy system, which considers the carbon flow coupling characteristics, takes the minimization of the park's overall operating cost as its objective function. The overall cost includes electricity purchase cost, equipment operation and maintenance cost, carbon emission cost, and wind and solar curtailment penalty cost. The constraints of the optimal scheduling model include power balance constraints, energy storage device status constraints, distributed power output constraints, flexible load adjustability constraints, and carbon flow rate constraints.

[0025] Among them, the carbon flow rate constraint is achieved by constructing a dynamic calculation model of node carbon potential. Specifically, based on the active power flow direction and carbon emission intensity of each node in the park, the real-time carbon potential of each node is calculated, and the upper limit constraint of node carbon potential is set, stipulating that the carbon potential value of each node at each time is less than or equal to the upper limit of node carbon potential.

[0026] The nodal carbon potential is calculated as follows: For any node n, obtain the set N of all neighboring nodes that inject power into that node. in And the power generation of the power generation units connected to node n itself. and its carbon emission intensity ;

[0027] The active power injected into each adjacent node Multiply by the carbon potential of the corresponding node Summing these values ​​and adding the carbon emission contribution from the node's own power generation, then dividing by the total active power injected into node n, yields the carbon potential of node n at time t. ;

[0028] The specific calculation process is as follows:

[0029] .

[0030] Furthermore, the process of solving the optimal scheduling model adopts an improved multi-objective particle swarm optimization algorithm. The improved multi-objective particle swarm optimization algorithm introduces adaptively adjusted inertial weights and selects the optimal solution set based on the Pareto front.

[0031] The adaptive update method for inertia weights is as follows:

[0032] In the k-th iteration, based on the current iteration number k, the maximum iteration number K, and the maximum value of the inertia weight... and minimum value And the change in global optimal fitness in the last two iterations. Calculate the inertia weight for the current iteration. ;

[0033] The inertia weights of the current iteration are updated using a linear decreasing strategy combined with a nonlinear adjustment term based on the change in fitness.

[0034] The specific calculation process is as follows:

[0035] ;

[0036] in, This is the adjustment coefficient;

[0037] The Pareto optimal solution set that satisfies all constraints is obtained by solving the improved multi-objective particle swarm optimization algorithm. Then, the compromise optimal solution is selected from the Pareto optimal solution set as the final scheduling scheme based on the fuzzy membership function.

[0038] Furthermore, the hierarchical distributed collaborative control includes a two-tier architecture of a park-level central controller and distributed edge computing terminals. The park-level central controller is responsible for executing the day-ahead scheduling plan and generating power allocation instructions for each sub-region based on the real-time correction plan. The distributed edge computing terminals are deployed at the distributed energy equipment, energy storage systems, and flexible load control nodes in each sub-region. They are responsible for collecting equipment status data in real time and executing autonomous control based on locally preset emergency control strategies when communication with the park-level central controller is interrupted. The emergency control strategies are obtained through offline training using reinforcement learning algorithms. These algorithms use local historical operating data as training samples and aim to minimize voltage deviation and maximize renewable energy consumption to generate a state-action mapping table. In island mode, the distributed edge computing terminals look up the state-action mapping table to execute control decisions.

[0039] Furthermore, the model-control closed-loop iterative optimization is achieved by constructing a dynamic evaluation mechanism for prediction deviation, which specifically includes: calculating the deviation between the predicted and actual power values ​​at each time scale in real time and generating a prediction deviation sequence;

[0040] Perform spectral analysis on the predicted deviation sequence to identify the main frequency components of the deviation and the corresponding sources of the deviation.

[0041] If the source of the deviation is attributed to a sudden meteorological event, the meteorological correction module of the hybrid forecasting model will be triggered to update the parameters.

[0042] If the source of the deviation is attributed to random load fluctuations, then the robust peer model of the optimization scheduling model is triggered to perform constraint relaxation adjustment.

[0043] If the deviation originates from equipment response delay, then update the response compensation parameters of the hierarchical distributed collaborative control.

[0044] The response compensation parameters are updated by calculating the actual response time of the most recent N control commands. With set response time Calculate the average deviation between the two, multiply it by the learning rate γ, and then combine it with the current response compensation parameter. Add them together to obtain the new response compensation parameters. ;

[0045] The specific calculation process is as follows:

[0046] .

[0047] Furthermore, it also includes steps for visualizing and tracing carbon flows in the park based on digital twins: constructing a digital twin model of the park's energy system, synchronously mapping multi-source data, power forecasting results, optimized scheduling schemes, and real-time control data to the digital twin model; in the digital twin model, constructing a carbon flow topology network of the park based on graph computing algorithms, tracking the source and destination of carbon flows at each node, and calculating the real-time carbon footprint of each energy-consuming unit; the real-time carbon footprint is calculated as follows: for any energy-consuming unit, from the initial time to the current time t, its power consumption is... Real-time carbon potential of access nodes Integrating the product of the two values ​​yields the cumulative carbon footprint of the energy-consuming unit. The calculation formula is:

[0048] ;

[0049] When the carbon potential of a node exceeds the warning threshold, the key path and main contributing equipment for carbon emission exceedance are located by reverse tracing through the carbon flow topology network, and carbon emission reduction auxiliary decision-making suggestions are generated.

[0050] A zero-carbon industrial park energy intelligent control system based on electricity forecasting, comprising:

[0051] The data acquisition module is used to acquire multi-source data from the park, including historical power load data, weather forecast data, renewable energy power generation data, and carbon emission monitoring data.

[0052] The hybrid forecasting module is used to perform multi-time-scale power forecasting based on multi-source data and through a pre-built hybrid forecasting model, generating ultra-short-term power forecasting sequences, short-term power forecasting sequences, and medium- and long-term power forecasting sequences.

[0053] The optimized scheduling model construction module is used to construct an optimized scheduling model for the park's energy system that takes into account carbon flow coupling characteristics, based on ultra-short-term power forecast sequences, short-term power forecast sequences, and medium- and long-term power forecast sequences.

[0054] The solution module is used to solve the optimization scheduling model and generate day-ahead scheduling plans and real-time correction plans that include multi-terminal coordination of source, grid, load and storage.

[0055] The hierarchical distributed control module is used to perform hierarchical distributed collaborative control of distributed energy equipment, energy storage systems and flexible loads in the park through an edge computing gateway based on the day-ahead scheduling plan and the real-time correction plan, and to collect control response data in real time.

[0056] The closed-loop optimization module is used to dynamically adjust the parameters of the hybrid prediction model and the constraints of the optimized scheduling model based on the deviation between the control response data and the corresponding power forecast value, forming a model-control closed-loop iterative optimization.

[0057] The present invention has the following advantages over the prior art:

[0058] By acquiring multi-source data from the industrial park and employing a hybrid prediction model that integrates physical mechanisms and data-driven approaches, we achieve multi-timescale power forecasting across ultra-short-term, short-term, and medium-to-long-term time scales. Ultra-short-term forecasts are corrected for real-time weather conditions, photovoltaic power fluctuations, and sudden weather events. Short-term forecasts utilize a bidirectional long short-term memory network incorporating an attention mechanism. Medium-to-long-term forecasts combine a seasonal autoregressive moving average model constrained by carbon emissions. This significantly improves the accuracy and adaptability of power forecasting, while also taking into account the park's carbon quota constraints, laying the foundation for zero-carbon management. The constructed optimal scheduling model for the park's energy system, considering carbon flow coupling characteristics, aims to minimize overall costs, encompassing electricity purchase, operation and maintenance, carbon emissions, and penalties for wind and solar curtailment. While considering costs, multiple constraints are set, including power balance, energy storage status, distributed power output, flexible load adjustability, and carbon flow rate based on dynamic calculation of node carbon potential. An improved multi-objective particle swarm optimization algorithm with adaptive inertia weights is used to solve the problem. Pareto front screening and fuzzy membership functions are used to select a compromise optimal solution. This generates a scientific day-ahead scheduling plan and real-time correction plan for multi-terminal collaborative operation of power generation, grid, load, and storage, effectively reducing the overall operating cost of the park, reducing carbon emissions, and improving the renewable energy absorption rate. A two-level hierarchical distributed collaborative control architecture is adopted, consisting of a park-level central controller and distributed edge computing terminals. The central controller coordinates the issuance of scheduling commands, while the edge computing terminals realize real-time data acquisition of equipment status and local... The autonomous control system employs an emergency control strategy trained offline using reinforcement learning, ensuring stable operation of the energy system even during communication interruptions and enhancing the reliability and robustness of the intelligent control system. By constructing a dynamic evaluation mechanism for prediction deviations, it achieves iterative optimization of the model-control closed loop. This allows for real-time analysis of the sources of deviation between predicted and actual power values, and targeted updates to hybrid prediction model parameters, adjustments to optimize scheduling model constraints, and updates to the response compensation parameters of hierarchical distributed collaborative control. This dynamic iterative optimization of the model and control continuously improves the adaptive capability and control accuracy of intelligent energy control. Furthermore, it integrates digital twin technology to construct a digital twin model of the park's energy system, enabling the integration of multi-source data, prediction results, scheduling schemes, and control... The synchronous mapping of data, based on graph computing algorithms, constructs a carbon flow topology network, which can track the source and destination of carbon flows in real time, calculate the cumulative carbon footprint of energy-consuming units, and trace the key paths and major contributing equipment of carbon emissions exceeding the threshold when the carbon potential of a node exceeds the threshold, and generate carbon emission reduction auxiliary decision-making suggestions. This enables the visualization and precise tracing of carbon flows in the park, helping the park achieve its zero-carbon goals. Overall, it realizes intelligent, refined, and hierarchical distributed management and control of park energy, forming a full-process intelligent control system of power forecasting, optimized scheduling, collaborative control, closed-loop optimization, and carbon flow tracing. This effectively ensures the safe, efficient, and low-carbon operation of the park's energy system and comprehensively improves the intelligent level of park energy management and zero-carbon control capabilities. Attached Figure Description

[0059] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0060] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0061] like Figure 1 As shown, the zero-carbon industrial park energy intelligent control method based on electricity forecasting includes the following steps:

[0062] Acquire multi-source data from the park, including historical power load data, weather forecast data, renewable energy power generation data, and carbon emission monitoring data;

[0063] Based on multi-source data, a pre-built hybrid prediction model is used to perform power forecasting at multiple time scales, generating ultra-short-term power forecasting sequences, short-term power forecasting sequences, and medium- and long-term power forecasting sequences.

[0064] Based on ultra-short-term power forecast sequences, short-term power forecast sequences, and medium- and long-term power forecast sequences, an optimal scheduling model for the park's energy system considering carbon flow coupling characteristics is constructed.

[0065] Solve the optimization scheduling model to generate a day-ahead scheduling plan and a real-time correction plan that includes multi-terminal coordination of source, grid, load and storage;

[0066] Based on the day-ahead scheduling plan and the real-time correction plan, the distributed energy equipment, energy storage system and flexible load in the park are controlled in a hierarchical distributed and coordinated manner through the edge computing gateway, and control response data is collected in real time.

[0067] Based on the deviation between the control response data and the corresponding power forecast values, the parameters of the hybrid forecasting model and the constraints of the optimized scheduling model are dynamically adjusted to form a model-control closed-loop iterative optimization.

[0068] Hybrid prediction models combine physical mechanism models with data-driven models, where:

[0069] The ultra-short-term power forecast sequence is generated using a similar day algorithm with real-time meteorological correction. Specifically, the measured meteorological data within a preset time period before the current moment is dynamically time-normalized and matched with historical meteorological data. Several historical days with the highest similarity are selected as a set of similar days. The power load data of the set of similar days is weighted and averaged to obtain the basic ultra-short-term forecast value. Based on the real-time output data of the photovoltaic inverter, the difference between the sliding average output value within the previous time window and the real-time output is calculated. This difference is used as the ultra-short-term photovoltaic power fluctuation correction amount and superimposed on the basic ultra-short-term forecast value to generate a preliminary ultra-short-term power forecast sequence.

[0070] The short-term power forecast sequence is generated by introducing a bidirectional long short-term memory network with an attention mechanism. The input features of the bidirectional long short-term memory network with the attention mechanism include weather forecast data, date type, historical load data, and park production plan data.

[0071] The medium- and long-term power forecast series is generated by combining a seasonal autoregressive moving average model with carbon emission constraints. The carbon emission constraints are obtained by allocating the total annual carbon quota of the park to each month according to the historical monthly load ratio, and then introducing the monthly allowable power consumption conversion factor into the model.

[0072] By combining physical mechanism models with data-driven models to construct a hybrid prediction model, the limitations of single models in power forecasting in terms of adaptability and accuracy are avoided. Differentiated prediction algorithms are designed for different time scales, including ultra-short-term, short-term, and medium-to-long-term forecasts, ensuring that prediction results at each scale accurately match the different needs of energy dispatching in the park. For ultra-short-term forecasts, a similar-day algorithm with real-time weather correction is used in conjunction with photovoltaic power fluctuation correction, balancing real-time and dynamic predictions to accurately support real-time energy dispatching in the park. For short-term forecasts, a bidirectional long short-term memory network with an attention mechanism is introduced to accurately capture the complex correlation between multi-dimensional input features such as weather and production plans and power load, improving short-term prediction accuracy and ensuring the scientific nature of day-ahead dispatching in the park. For medium-to-long-term forecasts, carbon emission constraints are integrated into a seasonal autoregressive moving average model, and the electricity consumption conversion factor is obtained through the monthly allocation of annual carbon quotas, ensuring that medium-to-long-term forecasts align with the long-term planning requirements for zero-carbon development in the park, achieving early integration of power forecasting and carbon control. High-precision, differentiated power forecasts across multiple time scales provide comprehensive, reliable, and practical data support for subsequent energy system optimization and dispatching in the park, ensuring the rationality and adaptability of subsequent dispatching schemes from the data source.

[0073] For example, generating ultra-short-term power forecast sequences:

[0074] The current time is set as 9:00. The hour before the current time (8:00-9:00) is taken as the preset time period. After dynamic time normalization and matching of the measured meteorological data such as sunshine and temperature during this period with historical meteorological data, the four historical days with the highest similarity are selected as the similar day set. The power load data of the four historical days in the ultra-short period of 9:00-9:30 are P1=220kW, P2=218kW, P3=223kW, and P4=219kW, respectively. Each similar day is assigned an equal weight wi=0.25 (i=1,2,3,4).

[0075] Calculate the basic ultra-short-term forecast values:

[0076] ;

[0077] Taking the 15 minutes before the current time (8:45-9:00) as the statistical time window for photovoltaic power output, the sliding average value of photovoltaic inverter output within this window is calculated. =95kW, inverter real-time output =90kW, calculate the correction amount for ultra-short-term photovoltaic power fluctuations:

[0078] ;

[0079] The core values ​​of the preliminary ultra-short-term power forecast sequence are generated by superimposing corrections:

[0080] ;

[0081] Based on this core value, and combined with the continuity of the time series, an ultra-short-term power forecast sequence [214kW, 215kW, 216kW, 215kW, 214kW, 213kW] (values ​​are taken at 5-minute intervals) is generated from 9:00 to 9:30.

[0082] Example of generating short-term power forecast sequences:

[0083] A bidirectional long short-term memory network (Attention-BiLSTM) with an attention mechanism was constructed to generate a 24-hour (short-term) power forecast sequence for the park. After the network was trained, the meteorological and operational data of the park on weekdays were selected as input features. The specific input features are as follows:

[0084] Meteorological forecast data: ambient temperature T=26℃, light intensity I=850W / ㎡, wind speed v=2.5m / s;

[0085] Date type: Weekdays are 1, rest days are 0, this example uses 1;

[0086] Historical load data: The average load Ph for the same period in the previous 3 days was 218kW;

[0087] Park production plan data: Daily production load coefficient kp=1.08 (meaning that the daily production load is 8% higher than that of a regular workday).

[0088] The network's attention layer assigns weights to each input feature, resulting in feature weights: temperature. Light intensity Wind speed Date type Historical load Production planning coefficient The sum of the weights is 1.

[0089] The Attention-BiLSTM network strengthens the influence of key features such as historical load and production plan through the attention layer, while weakening the interference of secondary features. After forward calculation by the network, it generates a short-term power forecast sequence for the park in the next 24 hours at 1-hour intervals. The predicted values ​​for the core periods are: morning peak (8:00-10:00) [232kW, 235kW], flat period (10:00-18:00) [220kW-228kW], evening peak (18:00-20:00) [230kW, 233kW], and off-peak period (23:00-6:00 the next day) [180kW-19kW]. This sequence accurately matches the production and power load patterns of the park on weekdays.

[0090] Example of generating medium- and long-term power forecast sequences:

[0091] A seasonal autoregressive moving average (SARIMA) model incorporating carbon emission constraints was used to generate a 12-month (medium- to long-term) power forecast series for the industrial park. The specific steps and calculations are as follows:

[0092] Determine the total annual carbon allowance for the park: Set the annual carbon emission allowance for the park. ;

[0093] Calculate the historical monthly load ratio: Statistically analyze the monthly power load data of the park over the past 3 years, calculate the ratio of each month's load to the total annual load, select the ratios for January to March as r1=0.07, r2=0.075, and r3=0.08 respectively, and allocate the ratios for the remaining months according to the actual load pattern. The sum of the ratios for all months is 1.

[0094] Allocate annual carbon allowances and calculate monthly electricity consumption conversion factors: Allocate the annual carbon allowances to each month according to the historical monthly load ratio to obtain the monthly carbon emission allowances. Combined with the carbon emission intensity of electricity consumption per unit in the park Calculate the monthly allowable electricity consumption conversion factor. ( (This is the historical average electricity consumption for that month).

[0095] Taking January as an example, Historical average electricity consumption in January ,but:

[0096] ;

[0097] A coefficient k1 = 1.05 indicates that the park's allowed electricity consumption in January can be increased by 5% compared to the historical average. If the coefficient is less than 1, it means that the corresponding proportion of electricity consumption needs to be reduced.

[0098] Introducing carbon emission constraints into the SARIMA model: adjusting the monthly electricity consumption coefficients As a constraint embedded in the SARIMA model's prediction formula, the model generates a medium- to long-term power forecast series for the park for the next 12 months by calculating using seasonal differencing and autoregressive moving averages, combined with constraint corrections based on the conversion factor. The predicted values ​​for January to March are as follows: The forecast results not only conform to the seasonal patterns of the park's historical load, but also meet the monthly carbon emission quota constraints, thus achieving the integration of medium- and long-term power forecasting and carbon management.

[0099] The preliminary ultra-short-term power forecast sequence still needs to be corrected by meteorological change response. Specifically, the meteorological change response mechanism is triggered when the rate of change of light intensity or temperature exceeds the set threshold during the preset time period by real-time monitoring of meteorological station data.

[0100] Calculate the change in light intensity over a preset time period prior to the current moment. and average light intensity and temperature change and average temperature T avg ;

[0101] rate of change of light intensity With temperature change rate Multiplying each factor by a preset weighting coefficient α and β and then summing the results yields the photovoltaic power output mutation correction coefficient M. pv , where α+β=1;

[0102] Multiply the preliminary ultra-short-term power forecast sequence by the photovoltaic power output mutation correction factor M. pv The corrected ultra-short-term power forecast sequence is obtained.

[0103] Photovoltaic power output mutation correction coefficient The calculation formula is:

[0104] ;

[0105] The above process adds a meteorological change response correction mechanism to the preliminary ultra-short-term power forecast sequence. It can monitor sudden changes in light intensity and temperature in real time and trigger targeted correction processes. By quantitatively calculating the photovoltaic output change correction coefficient, it can accurately adjust the forecast sequence, making up for the shortcomings of relying solely on similar day algorithms and photovoltaic power fluctuation correction to cope with sudden meteorological fluctuations. It effectively reduces the deviation of ultra-short-term power forecast caused by meteorological changes, and further improves the accuracy, robustness and dynamic adaptability of ultra-short-term power forecast. It makes the ultra-short-term forecast results more consistent with the actual meteorological changes and photovoltaic output of the park, providing a more accurate data source for the real-time correction and scheduling of the park's energy system, and ensuring the scientificity and effectiveness of the real-time scheduling strategy.

[0106] For example, in the previous unified scenario of ultra-short-term power forecasting: the current time is 9:00. After real-time weather correction and similar day algorithm + photovoltaic power fluctuation correction, the preliminary ultra-short-term power forecast sequence from 9:00 to 9:30 (5-minute interval) has been obtained.

[0107] [214kW, 215kW, 216kW, 215kW, 214kW, 213kW], with a core value of 215kW. A meteorological abrupt change response correction is now being performed on this sequence. The specific steps and calculations are as follows:

[0108] Set the threshold and weighting coefficient for monitoring meteorological changes:

[0109] The threshold values ​​for the rate of change of light intensity and the rate of change of temperature are set to 15% within a preset time period (8:00-9:00) and 10% respectively. The weighting coefficients for triggering the meteorological change response mechanism are α=0.6 (weight of the influence of light intensity on photovoltaic output) and β=0.4 (weight of the influence of temperature on photovoltaic output), satisfying α+β=1.

[0110] Collect and calculate parameters related to meteorological abrupt changes:

[0111] Real-time monitoring of weather station data from 8:00 to 9:00 AM allows for the calculation of changes in light intensity during this period. Average light intensity Temperature change Average temperature ;

[0112] Calculate the rate of change of light intensity: Temperature change rate: Both exceeded the set threshold, triggering the meteorological sudden change response correction mechanism.

[0113] Calculate the photovoltaic power output mutation correction factor M pv :

[0114] According to the formula Substitute the numerical values ​​into the calculation:

[0115] ;

[0116] The preliminary ultra-short-term power forecast series is revised as follows:

[0117] Multiply each value in the preliminary ultra-short-term power forecast sequence by the photovoltaic output mutation correction factor M. pv The corrected ultra-short-term power forecast sequence is obtained (the forecast values ​​are corrected synchronously due to the decrease in photovoltaic output caused by sudden weather changes), and the point-by-point calculation is as follows:

[0118] ;

[0119] The final revised 9:00-9:30 ultra-short-term power forecast sequence is [35.95kW, 36.12kW, 36.29kW, 36.12kW, 35.95kW, 35.78kW].

[0120] The optimal scheduling model for the park's energy system, considering the carbon flow coupling characteristics, takes the minimization of the park's overall operating cost as its objective function. The overall cost includes electricity purchase cost, equipment operation and maintenance cost, carbon emission cost, and wind and solar curtailment penalty cost. The constraints of the optimal scheduling model include power balance constraints, energy storage device status constraints, distributed power output constraints, flexible load adjustability constraints, and carbon flow rate constraints.

[0121] Among them, the carbon flow rate constraint is achieved by constructing a dynamic calculation model of node carbon potential. Specifically, based on the active power flow direction and carbon emission intensity of each node in the park, the real-time carbon potential of each node is calculated, and the upper limit constraint of node carbon potential is set, stipulating that the carbon potential value of each node at each time is less than or equal to the upper limit of node carbon potential.

[0122] The nodal carbon potential is calculated as follows: For any node n, obtain the set N of all neighboring nodes that inject power into that node. in And the power generation of the power generation units connected to node n itself. and its carbon emission intensity ;

[0123] The active power injected into each adjacent node Multiply by the carbon potential of the corresponding node Summing these values ​​and adding the carbon emission contribution from the node's own power generation, then dividing by the total active power injected into node n, yields the carbon potential of node n at time t. ;

[0124] The specific calculation process is as follows:

[0125] ;

[0126] The constructed optimal scheduling model for the park's energy system, considering carbon flow coupling characteristics, aims to minimize the overall operating cost of the park. It comprehensively covers electricity purchase costs, equipment operation and maintenance costs, carbon emission costs, and wind and solar curtailment penalties, avoiding energy imbalances caused by single-cost optimization. Simultaneously, it sets multiple constraints, including power balance, energy storage device status, distributed power output, flexible load adjustability, and carbon flow rate, to achieve dual constraint control of the park's energy system's physical operation and carbon emission management. The carbon flow rate constraint is implemented through a dynamic calculation model of node carbon potential, accurately quantifying the real-time carbon potential of each node in the park. By setting upper limits for carbon potential, this model achieves, for the first time, the coupled management and control of power flow and carbon flow in the park. This allows the scheduling scheme to not only meet the physical operating laws of energy equipment, but also to achieve precise control of carbon emissions at the node level. This effectively reduces the overall operating cost of the park while reducing carbon emissions and increasing the renewable energy absorption rate. The accurate calculation of node carbon potential can also provide core data support for subsequent carbon flow tracing and carbon emission exceedance location. This model has built a scientific framework for the optimized scheduling of the park's energy system that is economical, low-carbon, and feasible, ensuring the rationality and implementation of the multi-terminal collaborative scheduling scheme of source, grid, load, and storage.

[0127] This example continues the unified scenario of the aforementioned industrial park, selecting the core node n (the integrated power distribution station node of the industrial park) at time t=9:00 as the calculation object. Combining the objectives and constraints of the optimized scheduling model, it completes the calculation of node carbon potential, verification of carbon flow rate constraints, and demonstrates the core logic of model cost optimization. The specific parameters and calculations are as follows:

[0128] Node n has no connected power generation units, i.e. ;

[0129] The set of neighboring nodes to inject power into node n , ;

[0130] Injected power and nodal carbon potential at each adjacent node at time t=9:00: , (Photovoltaics is a clean energy source with extremely low carbon potential.) (The mains power grid includes thermal power, which has a high carbon potential.)

[0131] The model sets an upper limit on the carbon potential of node n. ;

[0132] Cost-related basic coefficients: electricity purchase cost c1 = 0.6 yuan / kWh, photovoltaic equipment operation and maintenance cost c2 = 0.02 yuan / kWh, carbon emission cost c3 = 0.1 yuan / kgCO2, and wind and solar curtailment penalty cost c4 = 0.8 yuan / kWh.

[0133] Calculate the carbon potential of node n at time t. :

[0134] Based on the formula for calculating nodal carbon potential:

[0135] ;

[0136] Substitute the numerical values ​​into the calculation:

[0137] ;

[0138] Carbon flow rate constraint verification and preliminary scheduling adjustments:

[0139] Calculated This triggers a carbon flow rate constraint, and the optimized scheduling model will adjust the scheduling according to the goal of minimizing overall cost: increasing the output of photovoltaic power plants to... Reduce grid connection power to With all other parameters unchanged, recalculate the nodal carbon potential:

[0140]

[0141] at this time This satisfies the carbon flow rate constraint.

[0142] Verify the overall cost optimization effect after scheduling adjustment:

[0143] Calculate the comprehensive cost at time t before and after adjustment (only the core cost item is calculated, unit: yuan). The comprehensive cost formula is simplified to:

[0144] ;

[0145] Total cost before adjustment:

[0146] ;

[0147] Adjusted total cost:

[0148] ;

[0149] The adjusted overall cost is reduced by RMB 6.3805 (84.4755 - 78.095) compared to the previous cost, and there is no penalty cost for curtailment of wind and solar power. At the same time, it meets the power balance constraint (the total injected power is 215kW, matching the load demand of the park at time t) and the distributed power output constraint (the photovoltaic output is within its rated output range), thus achieving the dual goals of minimizing costs and satisfying multiple constraints.

[0150] The process of solving the optimal scheduling model adopts an improved multi-objective particle swarm optimization algorithm. The improved multi-objective particle swarm optimization algorithm introduces adaptively adjusted inertia weights and selects the optimal solution set based on the Pareto front.

[0151] The adaptive update method for inertia weights is as follows:

[0152] In the k-th iteration, based on the current iteration number k, the maximum iteration number K, and the maximum value of the inertia weight... and minimum value And the change in global optimal fitness in the last two iterations. Calculate the inertia weight for the current iteration. ;

[0153] The inertia weights of the current iteration are updated using a linear decreasing strategy combined with a nonlinear adjustment term based on the change in fitness.

[0154] The specific calculation process is as follows:

[0155] ;

[0156] in, This is the adjustment coefficient;

[0157] The Pareto optimal solution set that satisfies all constraints is obtained by solving the improved multi-objective particle swarm optimization algorithm. Then, the compromise optimal solution is selected from the Pareto optimal solution set as the final scheduling scheme based on the fuzzy membership function.

[0158] An improved multi-objective particle swarm optimization (PSO) algorithm incorporating adaptive inertia weights is employed to solve the optimal scheduling model. This addresses the problems of imbalance between global exploration and local exploitation capabilities, premature convergence, and low solution accuracy caused by the fixed inertia weights in traditional PSO algorithms. By dynamically updating the inertia weights through a linear decreasing strategy combined with a nonlinear adjustment term based on fitness changes, the algorithm maintains a larger weight in the early stages of iteration to enhance global search capabilities, while reducing the weight in the later stages to improve local exploitation accuracy. Furthermore, real-time adjustments based on the global optimal fitness change allow the algorithm to adapt to the solution characteristics of the scheduling model, improving both convergence speed and solution accuracy. Then, by using Pareto front screening, the optimal solution set that satisfies multiple objectives is obtained, avoiding the shortcomings of a single optimal solution that cannot take into account the multi-dimensional needs of the park's comprehensive cost, carbon emissions, and renewable energy absorption rate. Finally, based on the fuzzy membership function, a compromise optimal solution is selected from the Pareto optimal solution set, so that the generated scheduling scheme not only satisfies all the constraints of the model, but also takes into account economy, low carbon emissions, and engineering feasibility. Compared with traditional solution algorithms, this method can solve the park energy system optimization scheduling model that considers carbon flow coupling characteristics more efficiently and accurately, and the resulting multi-terminal coordinated scheduling scheme of source, grid, load, and storage is more in line with the actual operation needs of the park.

[0159] This example continues the unified scenario of the aforementioned park, addressing the energy dispatch optimization problem at time t=9:00 in the park. An improved multi-objective particle swarm optimization algorithm is used to solve the problem, with multiple optimization objectives including minimizing overall cost, minimizing node carbon potential, and maximizing photovoltaic absorption rate. The algorithm completes adaptive inertia weight calculation, iterative optimization, and selection of the optimal solution. Specific parameters and calculations are as follows:

[0160] Setting the basic parameters for the improved multi-objective particle swarm optimization algorithm:

[0161] To address the scheduling problem in this park, the core parameters of the algorithm are set as follows: maximum number of iterations K=100, current number of iterations k=50, and maximum value of inertia weight. Minimum inertial weight adjustment coefficient The change in global optimal fitness between the last two iterations (A smaller fitness value indicates a better solution.) This indicates that the fitness of this iteration has improved compared to the previous one.

[0162] Calculate the adaptive inertia weights for the current iteration. :

[0163] According to the adaptive inertia weight update formula:

[0164] ;

[0165] Substitute the values ​​and calculate step by step:

[0166] Calculation of linearly decreasing terms: ;

[0167] Calculation of nonlinear adjustment term: ;

[0168] Total inertia weight calculation: .

[0169] This iteration It falls within the range of 0.4 to 0.9, balancing the algorithm's global exploration and local development capabilities, and is suitable for the optimization needs in the middle of the iteration process.

[0170] Algorithm Iterative Optimization and Pareto Optimal Solution Set Generation:

[0171] Based on the overall cost of the park (C) and the carbon potential at each node The photovoltaic grid integration rate η is one of the three optimization objectives (C and...). (The smaller the better, the larger the better), constraints such as power balance, carbon flow rate, and distributed power output are used as optimization constraints for the algorithm. Through iterative calculation using an improved multi-objective particle swarm optimization algorithm, when the algorithm converges to k=100 times, a Pareto optimal solution set satisfying all constraints is obtained. Three typical non-dominated solutions in the solution set are selected as candidate solutions, and the objective values ​​of each candidate solution are as follows:

[0172] Solution 1: C1 = 77.5 yuan ;

[0173] Solution 2: C2 = 78.1 yuan ;

[0174] Solution 3: C3 = 80.2 yuan .

[0175] Selecting the optimal compromise solution based on fuzzy membership functions:

[0176] Normalize the objective values ​​in the Pareto optimal solution set and construct fuzzy membership functions. (j is the target number, i is the candidate solution number). The closer the membership value is to 1, the better the solution performs on the corresponding target. Then calculate the average membership of each candidate solution. The solution with the largest average membership is the compromise optimal solution.

[0177] Set target weights: Overall cost weights Node carbon potential weight Photovoltaic grid integration rate weight (The weighted sum is 1, which aligns with the park's scheduling requirements of "prioritizing economic efficiency while taking into account low carbon emissions and energy consumption").

[0178] Calculate the fuzzy membership degree (after normalization) and weighted membership degree of each candidate solution:

[0179] Solution 1: Weighted membership ;

[0180] Solution 2: Weighted membership degree ;

[0181] Solution 3: Weighted membership degree ;

[0182] Compromise optimal solution selection: Weighted membership degree of solution 2 Since the maximum value is 100kW, solution 2 is selected as the final scheduling scheme. The corresponding scheduling instruction is 100kW output of photovoltaic power station, 115kW grid access power, and no charging or discharging of energy storage system. This scheme takes into account the multi-dimensional needs of park comprehensive cost, carbon emission control and photovoltaic absorption rate, and fully meets all the constraints of the model. It is highly compatible with the actual operation scenario of the park.

[0183] The hierarchical distributed collaborative control system comprises a two-tier architecture: a park-level central controller and distributed edge computing terminals. The park-level central controller is responsible for executing the day-ahead scheduling plan and generating power allocation instructions for each sub-region based on the real-time correction plan. The distributed edge computing terminals are deployed at distributed energy devices, energy storage systems, and flexible load control nodes in each sub-region. They are responsible for collecting real-time device status data and executing autonomous control based on locally pre-set emergency control strategies when communication with the park-level central controller is interrupted. The emergency control strategies are obtained through offline training using reinforcement learning algorithms. These algorithms use local historical operating data as training samples and aim to minimize voltage deviation and maximize renewable energy consumption. They generate a state-action mapping table, and the distributed edge computing terminals execute control decisions by looking up the state-action mapping table in islanded mode.

[0184] A two-tiered, distributed, collaborative control architecture combining a park-level central controller and distributed edge computing terminals organically integrates overall park energy dispatch planning with local fine-grained control. The central controller coordinates the execution of day-ahead dispatch plans and generates real-time power allocation instructions, ensuring the rationality of global dispatching across multiple sources (sources, grid, load, and storage). Distributed edge computing terminals are deployed at each device control node, enabling real-time acquisition of device status data to ensure control instructions align with actual device operating conditions. This significantly reduces the computational load on the central controller, improving the overall system control response speed. Furthermore, when communication with the central controller is interrupted, the edge terminals can execute local autonomous control based on offline reinforcement learning training, solving the problem of loss of control due to communication interruption in traditional centralized control. The reinforcement learning algorithm trains and generates a state-action mapping table with the objectives of minimizing voltage deviation and maximizing renewable energy consumption. This eliminates the need for complex real-time calculations in islanded mode control decisions, enabling rapid and accurate issuance of control orders. This effectively ensures the dual operational stability of the park's energy system under both normal and islanded operating conditions, enhancing the robustness, fault tolerance, and engineering feasibility of the intelligent control system. It provides a reliable control execution system for the efficient implementation of dispatch plans.

[0185] Continuing with the aforementioned unified scenario for the industrial park, and based on the park's scheduling plan at t=9:00 (100kW photovoltaic output, 115kW grid connection, 215kW comprehensive load, and no charging or discharging of the energy storage system), the park is divided into three sub-regions: a photovoltaic power station sub-region (terminal 1), an energy storage system sub-region (terminal 2), and a comprehensive load sub-region (terminal 3). Distributed edge computing terminals are deployed in each sub-region. The park's energy management center has a central controller. The edge terminals communicate with the central controller via industrial Ethernet. Emergency control strategies trained offline through reinforcement learning are implemented based on node voltage deviation rates. Minimum photovoltaic absorption rate The optimization objective is to maximize the control process, and the specific control flow and islanded operating condition verification are as follows:

[0186] Two-level hierarchical distributed collaborative control under normal operating conditions:

[0187] Based on the obtained optimal scheduling scheme, the park-level central controller generates power allocation instructions for each sub-region: Terminal 1 (photovoltaic) maintains a rated output of 100kW, and Terminal 2 (energy storage) charges and discharges power. Terminal 3 (comprehensive load) has a stable power load of 215kW;

[0188] Each distributed edge computing terminal collects local device status data in real time: Terminal 1 collects photovoltaic inverter output of 100kW and DC side voltage of 800V; Terminal 2 collects energy storage SOC value of 60% and AC side voltage of 380V; Terminal 3 collects load side voltage of 380V and real-time load of 215kW.

[0189] Each terminal uploads its device status data to the central controller in real time. The central controller compares the actual status with the scheduling instructions. If there is no discrepancy, the original instructions are maintained, thus achieving coordination between global planning and local real-time monitoring. Under this operating condition, the device status and instructions are perfectly matched, and the system operates stably.

[0190] Local autonomous emergency control in isolated conditions:

[0191] The simulation assumes an isolated operating condition where communication between terminal 1 and the central controller is interrupted. In this case, the photovoltaic power station sub-area is controlled autonomously only by terminal 1. Terminal 1 has a built-in state-action mapping table trained offline using reinforcement learning. Its core state-action matching rules have been solidified through offline training. The specific execution process and calculation are as follows:

[0192] Setting core parameters and determining state variables:

[0193] Rated voltage on the low-voltage side of the park Voltage deviation rate threshold Rated output of photovoltaic power station In isolated operating conditions, sudden fluctuations in local small loads to There is a surplus of 20kW of photovoltaic power output;

[0194] Terminal 1 collects local node voltage in real time Calculate the voltage deviation rate:

[0195] ;

[0196] The current state variable is determined as follows: Exceeding the 5% threshold, the remaining output of photovoltaic power... In the 10kW to 20kW range, terminal 1 matches the corresponding control action according to the built-in state-action mapping table.

[0197] State-action mapping matching and control action issuance:

[0198] The state-action mapping table built into Terminal 1 is trained offline through reinforcement learning, which solidifies the optimal control actions corresponding to different state combinations. The rule that matches the current working condition is: when the node voltage deviation rate is >5% and the remaining photovoltaic output is in the range of 10kW to 20kW, a 18kW charging command is issued to the energy storage terminal 2 and a 2kW flexible load boosting command is issued to the integrated load terminal 3. Terminal 1 issues this coordinated control action to Terminal 2 and Terminal 3 simultaneously through the local communication link in the park, without the need for the central controller, thus realizing local autonomous decision-making.

[0199] Verification of the effect after the control action is executed:

[0200] Terminals 2 and 3 execute the control actions immediately upon receiving them. Terminal 1 collects the equipment operation data in real time after execution and calculates key indicators to verify the achievement of optimization goals.

[0201] Calculation of photovoltaic grid integration rate: With no curtailment of solar power output throughout this project, the photovoltaic grid integration rate is:

[0202] ;

[0203] To achieve the optimization goal of full photovoltaic power consumption;

[0204] Voltage deviation rate calculation: After executing the control action, the local node voltage drops to... The new voltage deviation rate;

[0205] ;

[0206] The voltage deviation rate decreased from 5.26% to 0.26%, far below the 5% threshold, achieving the optimization goal of minimizing voltage deviation;

[0207] Power balance verification: After the control is executed, the 100kW photovoltaic output fully matches the total absorption power of 80kW base load + 2kW flexible load + 18kW energy storage charging, meeting the power balance constraints, and the sub-area power system of the park maintains stable operation.

[0208] In this isolated operating condition, the distributed edge computing terminal quickly completed autonomous control decisions and issued multi-terminal collaborative actions through a pre-set reinforcement learning state-action mapping table. This not only accurately achieved the preset optimization goals of minimizing voltage deviation and fully absorbing photovoltaic power, but also ensured the stability of the power system in the sub-area of ​​the park. This verified the high reliability and fault tolerance of the two-level hierarchical distributed collaborative control architecture under abnormal operating conditions, and also demonstrated the efficient support of this control method for the implementation of scheduling plans.

[0209] Model-control closed-loop iterative optimization is achieved by constructing a dynamic evaluation mechanism for prediction deviation, which specifically includes: calculating the deviation between the predicted and actual power values ​​at each time scale in real time and generating a prediction deviation sequence;

[0210] Perform spectral analysis on the predicted deviation sequence to identify the main frequency components of the deviation and the corresponding sources of the deviation.

[0211] If the source of the deviation is attributed to a sudden meteorological event, the meteorological correction module of the hybrid forecasting model will be triggered to update the parameters.

[0212] If the source of the deviation is attributed to random load fluctuations, then the robust peer model of the optimization scheduling model is triggered to perform constraint relaxation adjustment.

[0213] If the deviation originates from equipment response delay, then update the response compensation parameters of the hierarchical distributed collaborative control.

[0214] The response compensation parameters are updated by calculating the actual response time of the most recent N control commands. With set response time Calculate the average deviation between the two, multiply it by the learning rate γ, and then combine it with the current response compensation parameter. Add them together to obtain the new response compensation parameters. ;

[0215] The specific calculation process is as follows:

[0216] ;

[0217] By constructing a dynamic evaluation mechanism for prediction deviations to achieve iterative optimization of the model-control closed loop, the system can calculate the deviation between predicted and actual power values ​​at various time scales in real time and generate a deviation sequence. Through spectrum analysis, the source of deviations is accurately identified, and differentiated optimization and adjustment strategies are adopted for different sources of deviation, such as sudden weather changes, random load fluctuations, and equipment response delays. This achieves precise targeted optimization of the hybrid prediction model, optimized scheduling model, and hierarchical distributed collaborative control, solving the problems of model-control disconnect and fixed parameters leading to decreased adaptability over time in traditional energy intelligent control systems. In particular, the equipment response compensation parameters are quantitatively updated by statistically analyzing historical response deviations and combining them with the learning rate, making the control compensation more in line with the actual operating characteristics of the equipment. This closed-loop iterative optimization mechanism enables the prediction, scheduling, and control systems of the park's energy intelligent control to form a dynamic and self-optimizing whole, continuously improving the adaptive capability, prediction accuracy, and control response efficiency of the intelligent control system. This allows the intelligent control system to dynamically adapt to real-time changes in the park's weather, load, and equipment operating status, ensuring the long-term accurate implementation of the park's energy scheduling and control strategies, and further improving the stability, economy, and low-carbon operation of the energy system.

[0218] This example continues the unified scenario of the aforementioned park, based on the control commands of 100kW photovoltaic output, 18kW energy storage charging, and 2kW flexible load boost at t=9:00 in the park. It completes the entire process of model-control closed-loop iterative optimization, targeting the sources of equipment response delay deviations that occur during the operation of the intelligent control system. The focus is on updating and verifying the effect of hierarchical distributed collaborative control response compensation parameters. Specific parameter settings and calculations are as follows:

[0219] Deviation identification and judgment:

[0220] Real-time acquisition of control response data from various devices in the park; calculation of the deviation between ultra-short-term power forecasts and actual values; generation of a deviation sequence; and spectral analysis to determine that the source of the deviation is equipment response delay, specifically manifested as the actual response time of the energy storage system and flexible loads executing control commands exceeding the set response time, necessitating an update to the response compensation parameters. Set the core update parameters as follows: most recent statistics count N=5, learning rate γ=0.2, and current response compensation parameter. Control commands set response time (All device commands have a uniform response time setting).

[0221] Historical control command response time statistics:

[0222] Analyze the actual response time of the device for the five most recent control commands of the same type. The time intervals were 1.2s, 1.3s, 1.1s, 1.4s, and 1.2s, respectively. The deviation between the response time and the set response time was calculated for each interval. ,get .

[0223] Calculate the new response compensation parameters using the formula. :

[0224] Update the formula based on the response compensation parameters:

[0225] ;

[0226] Step-by-step numerical calculation:

[0227] Calculate the sum of the deviations: ;

[0228] Calculate the average deviation: ;

[0229] Calculate the learning rate correction term: ;

[0230] Calculate the new compensation parameters: .

[0231] Finally, the updated response compensation parameters are obtained. This parameter will be synchronously sent to each distributed edge computing terminal for compensation of the response time of subsequent control commands.

[0232] New response compensation parameters By embedding a hierarchical distributed collaborative control system, the same type of control command (100kW photovoltaic output, 18kW energy storage charging, and 2kW flexible load boost) was issued to the park again. The actual response time of the equipment was collected in real time, and the actual response time of the energy storage system and the flexible load was found to be 1.05s, with a deviation of only 0.05s from the set response time of 1s. This is a significant reduction compared to the average deviation of 0.24s before the update. At the same time, the deviation rate between the power prediction value and the actual value decreased from 3.2% before the update to 0.8%. This verifies that after closed-loop iterative optimization, the accuracy of equipment control response has been significantly improved, the prediction deviation has been effectively reduced, and the adaptive capability and control accuracy of the intelligent control system have been effectively improved.

[0233] If the spectral analysis determines that the source of the deviation is a sudden meteorological event, the meteorological correction module of the hybrid forecasting model will be directly triggered to update the core parameters such as the weighting coefficients α and β of illumination and temperature, thereby improving the adaptability of ultra-short-term forecasts to sudden meteorological events. If the source of the deviation is determined to be random load fluctuations, the robust equivalence model of the optimized scheduling model will be triggered to appropriately relax the adjustable constraints of flexible loads and the constraints of energy storage output, thereby expanding the optimization range of the scheduling model and making the scheduling scheme more adaptable to the random variation characteristics of the load, achieving targeted optimization for different sources of deviation.

[0234] It also includes steps for visualizing and tracing carbon flows in the park based on digital twins: constructing a digital twin model of the park's energy system, synchronously mapping multi-source data, power forecasting results, optimized scheduling schemes, and real-time control data to the digital twin model; in the digital twin model, constructing a carbon flow topology network of the park based on graph computing algorithms, tracking the source and destination of carbon flows at each node, and calculating the real-time carbon footprint of each energy-consuming unit; the real-time carbon footprint is calculated as follows: for any energy-consuming unit, from the initial time to the current time t, its power consumption is... Real-time carbon potential of access nodes Integrating the product of the two values ​​yields the cumulative carbon footprint of the energy-consuming unit. The calculation formula is:

[0235] ;

[0236] When the carbon potential of a node exceeds the warning threshold, the key path and main contributing equipment for carbon emission exceedance are located by reverse tracing through the carbon flow topology network, and carbon emission reduction auxiliary decision-making suggestions are generated.

[0237] By constructing a digital twin model of the park's energy system, synchronous mapping of multi-source data, power forecast results, optimized scheduling schemes, and real-time control data has been achieved, enabling precise linkage between the park's energy physical system and the digital model. The carbon flow topology network built based on graph computing algorithms can accurately track the source and destination of carbon flows at each node in the park. At the same time, quantitative statistics of the real-time accumulated carbon footprint of each energy-consuming unit are achieved through integral calculation, solving the problems of unintuitive carbon flow monitoring and inaccurate carbon footprint accounting in traditional carbon management. When the carbon potential of a node exceeds the warning threshold, the carbon flow topology network can be used to quickly trace back to locate the key path and main contributing equipment for carbon emission exceeding the standard, avoiding the blindness of carbon emission reduction investigation. It can also generate targeted carbon emission reduction auxiliary decision-making suggestions, providing accurate data support and action guidance for the park's carbon management, realizing visualized, refined, and traceable management of carbon flows in the park, improving the whole-process closed loop of zero-carbon park energy intelligent control, and transforming carbon management from passive monitoring to proactive tracing and precise policy implementation, greatly improving the intelligence and operability of zero-carbon management in the park.

[0238] Continuing with the unified scenario of the aforementioned industrial park, and taking the integrated substation node n as the core monitoring node, and the industrial park production workshop as a typical energy-consuming unit (the unit is uniquely connected to node n), a digital twin model of the industrial park's energy system is constructed. This model synchronizes previous multi-source data, scheduling schemes, and control data. A carbon flow topology network is built based on graph computing algorithms (covering photovoltaic power station node i1, grid connection node i2, integrated substation node n, and the production workshop energy-consuming unit). The cumulative carbon footprint of the production workshop is calculated, and source tracing and emission reduction recommendations are generated after the node's carbon potential exceeds the warning threshold. Specific parameters and calculations are as follows:

[0239] Digital twin model construction and data synchronization:

[0240] Construct a digital twin model that maps 1:1 to the physical energy system of the park, and store the real-time power data of nodes i1, i2, and n. Carbon potential data () Data such as power consumption in the production workshop, optimization scheduling and adjustment instructions, and equipment control response data are fully synchronized to the digital twin model. The model updates data in real time at a 1-minute granularity, realizing dynamic linkage between the physical system and the digital model. At the same time, a carbon flow topology network is built based on graph computing algorithms, defining the carbon flow propagation path as: photovoltaic power station i1 → integrated distribution station n → production workshop, grid access i2 → integrated distribution station n → production workshop, clarifying the correlation between carbon flow transmission power and carbon potential for each path.

[0241] Set the basic parameters for carbon footprint calculation:

[0242] Selecting the initial time The carbon footprint statistics period is from t=9:00 to the present, lasting 1 hour. The production workshop, as an energy-consuming unit, is connected to node n of the integrated power distribution station. Its power consumption remains stable during the period from 8:00 to 9:00. (τ∈[8:00,9:00]); the carbon potential of node n before scheduling adjustment during this period. (8:00-8:30), adjusted carbon potential (8:30-9:00), the warning threshold is set to .

[0243] Calculate the cumulative carbon footprint of the production workshop ;

[0244] According to the real-time carbon footprint calculation formula:

[0245] ;

[0246] Because power and carbon potential remain constant in two segments during the statistical period, the definite integral is calculated by summing the definite integrals of two sub-intervals. The integral result corresponds to the total carbon footprint (unit: kgCO2). The specific calculation steps are as follows:

[0247] Split the integration interval: ;

[0248] Calculate the integral value for the subinterval, converting the time unit to hours, with 0.5 hours as the upper limit for integration:

[0249] ;

[0250] ;

[0251] Summing yields the cumulative carbon footprint:

[0252] ;

[0253] The cumulative carbon footprint of the production workshop during the period from 8:00 to 9:00 is 73.7025 kg CO2, and this value is displayed in real time in the carbon management module of the energy-consuming unit of the digital twin model.

[0254] Reverse source tracing and emission reduction recommendations for node carbon potential exceeding the threshold:

[0255] In digital twin model monitoring, the carbon potential of node n during the 8:00-8:30 time period. This triggers a carbon exceedance warning, prompting reverse tracing based on carbon flow topology networks:

[0256] The carbon flow input sources for node n are the photovoltaic power plant i1 and the grid connection node i2, where the carbon potential of i1 is... (Clean energy, extremely low carbon flow contribution), i2 carbon potential (High carbon potential power source), and during this period The excessively high proportion of grid-connected power is the core reason for the excessive carbon potential at node n.

[0257] The carbon flow contribution ratio of each input node was calculated using a carbon flow topology network. Carbon flow contribution = injected power × node carbon potential. The carbon flow contribution of i1 = 90 × 0.02 = 1.8 kg CO2 / kWh, and the carbon flow contribution of i2 = 125 × 0.6 = 75 kg CO2 / kWh. The carbon flow contribution ratio of i2 = 75 / (1.8 + 75) × 100% ≈ 97.67%. It was determined that the grid access node i2 is the main contributing device for carbon emission exceedance, and high-power grid access is the key path for exceedance.

[0258] Based on the source tracing results and the operating characteristics of energy equipment in the park, the digital twin model generates targeted carbon emission reduction auxiliary decision-making suggestions: increase the output of photovoltaic power station i1 to 100kW, simultaneously reduce the output of grid access node i2 to 115kW, and maintain the total injected power of 215kW to match the load of the production workshop. This suggestion is consistent with the previous optimized scheduling adjustment strategy. After execution, the carbon potential of node n drops to 0.330kgCO2 / kWh, which is lower than the warning threshold, thus achieving the carbon emission reduction target.

[0259] The effects of carbon flow visualization and traceability are demonstrated as follows:

[0260] The digital twin model visualizes the carbon flow topology network, the carbon potential of each node, the carbon footprint of energy-consuming units, carbon exceedance paths, and emission reduction recommendations. Staff can directly view the real-time transmission status of the carbon flow and the quantitative data of carbon emissions of each unit in the model. Carbon exceedance issues can be quickly located without manual on-site inspection, which greatly improves the efficiency and accuracy of carbon management in the park. At the same time, the accumulated carbon footprint data also provides accurate quantitative basis for the park's annual and monthly carbon quota allocation and carbon assessment of energy-consuming units.

[0261] A zero-carbon industrial park energy intelligent control system based on electricity forecasting, comprising:

[0262] The data acquisition module is used to acquire multi-source data from the park, including historical power load data, weather forecast data, renewable energy power generation data, and carbon emission monitoring data.

[0263] The hybrid forecasting module is used to perform multi-time-scale power forecasting based on multi-source data and through a pre-built hybrid forecasting model, generating ultra-short-term power forecasting sequences, short-term power forecasting sequences, and medium- and long-term power forecasting sequences.

[0264] The optimized scheduling model construction module is used to construct an optimized scheduling model for the park's energy system that takes into account carbon flow coupling characteristics, based on ultra-short-term power forecast sequences, short-term power forecast sequences, and medium- and long-term power forecast sequences.

[0265] The solution module is used to solve the optimization scheduling model and generate day-ahead scheduling plans and real-time correction plans that include multi-terminal coordination of source, grid, load and storage.

[0266] The hierarchical distributed control module is used to perform hierarchical distributed collaborative control of distributed energy equipment, energy storage systems and flexible loads in the park through an edge computing gateway based on the day-ahead scheduling plan and the real-time correction plan, and to collect control response data in real time.

[0267] The closed-loop optimization module is used to dynamically adjust the parameters of the hybrid prediction model and the constraints of the optimized scheduling model based on the deviation between the control response data and the corresponding power forecast value, forming a model-control closed-loop iterative optimization.

[0268] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A zero-carbon industrial park energy intelligent control method based on power forecasting, characterized in that, Includes the following steps: Acquire multi-source data from the park, including historical power load data, weather forecast data, renewable energy power generation data, and carbon emission monitoring data; Based on multi-source data, a pre-built hybrid prediction model is used to perform power forecasting at multiple time scales, generating ultra-short-term power forecasting sequences, short-term power forecasting sequences, and medium- and long-term power forecasting sequences. Based on ultra-short-term power forecast sequences, short-term power forecast sequences, and medium- and long-term power forecast sequences, an optimal scheduling model for the park's energy system considering carbon flow coupling characteristics is constructed. Solve the optimization scheduling model to generate a day-ahead scheduling plan and a real-time correction plan that includes multi-terminal coordination of source, grid, load and storage; Based on the day-ahead scheduling plan and the real-time correction plan, the distributed energy equipment, energy storage system and flexible load in the park are controlled in a hierarchical distributed and coordinated manner through the edge computing gateway, and control response data is collected in real time. Based on the deviation between the control response data and the corresponding power forecast value, the parameters of the hybrid forecast model and the constraints of the optimized scheduling model are dynamically adjusted to form a model-control closed-loop iterative optimization. Hybrid prediction models combine physical mechanism models with data-driven models, where: The ultra-short-term power forecast sequence is generated using a similar day algorithm with real-time meteorological correction. Specifically, the measured meteorological data within a preset time period before the current moment is dynamically time-normalized and matched with historical meteorological data. Several historical days with the highest similarity are selected as a set of similar days. The power load data of the set of similar days is weighted and averaged to obtain the basic ultra-short-term forecast value. Based on the real-time output data of the photovoltaic inverter, the difference between the sliding average output value within the previous time window and the real-time output is calculated. This difference is used as the ultra-short-term photovoltaic power fluctuation correction amount and superimposed on the basic ultra-short-term forecast value to generate a preliminary ultra-short-term power forecast sequence. The short-term power forecast sequence is generated by introducing a bidirectional long short-term memory network with an attention mechanism. The input features of the bidirectional long short-term memory network with the attention mechanism include weather forecast data, date type, historical load data, and park production plan data. The medium- and long-term power forecast series is generated by combining a seasonal autoregressive moving average model with carbon emission constraints. The carbon emission constraints are obtained by allocating the total annual carbon quota of the park to each month according to the historical monthly load ratio, and then introducing the monthly allowable power consumption conversion factor into the model. The optimal scheduling model for the park's energy system, considering the carbon flow coupling characteristics, takes the minimization of the park's overall operating cost as its objective function. The overall cost includes electricity purchase cost, equipment operation and maintenance cost, carbon emission cost, and wind and solar curtailment penalty cost. The constraints of the optimal scheduling model include power balance constraints, energy storage device status constraints, distributed power output constraints, flexible load adjustability constraints, and carbon flow rate constraints. Among them, the carbon flow rate constraint is achieved by constructing a dynamic calculation model of node carbon potential. Specifically, based on the active power flow direction and carbon emission intensity of each node in the park, the real-time carbon potential of each node is calculated, and the upper limit constraint of node carbon potential is set, stipulating that the carbon potential value of each node at each time is less than or equal to the upper limit of node carbon potential. The nodal carbon potential is calculated as follows: For any node n, obtain the set N of all neighboring nodes that inject power into that node. in And the power generation of the power generation units connected to node n itself. and its carbon emission intensity ; The active power injected into each adjacent node Multiply by the carbon potential of the corresponding node Summing these values ​​and adding the carbon emission contribution from the node's own power generation, then dividing by the total active power injected into node n, yields the carbon potential of node n at time t. .

2. The zero-carbon industrial park energy intelligent control method based on power forecasting according to claim 1, characterized in that: The preliminary ultra-short-term power forecast sequence still needs to be corrected by meteorological change response. Specifically, the meteorological change response mechanism is triggered when the rate of change of light intensity or temperature exceeds the set threshold during the preset time period by real-time monitoring of meteorological station data. Calculate the change in light intensity over a preset time period prior to the current moment. and average light intensity and temperature change and average temperature T avg ; rate of change of light intensity With temperature change rate Multiplying each factor by a preset weighting coefficient α and β and then summing the results yields the photovoltaic power output mutation correction coefficient M. pv , where α+β=1; Multiply the preliminary ultra-short-term power forecast sequence by the photovoltaic power output mutation correction factor M. pv The corrected ultra-short-term power forecast sequence was obtained.

3. The zero-carbon park energy intelligent control method based on power forecasting according to claim 2, characterized in that: The process of solving the optimal scheduling model adopts an improved multi-objective particle swarm optimization algorithm. The improved multi-objective particle swarm optimization algorithm introduces adaptively adjusted inertia weights and selects the optimal solution set based on the Pareto front. The adaptive update method for inertia weights is as follows: In the k-th iteration, based on the current iteration number k, the maximum iteration number K, and the maximum value of the inertia weight... and minimum value And the change in global optimal fitness in the last two iterations. Calculate the inertia weight for the current iteration; The inertia weights of the current iteration are updated using a linear decreasing strategy combined with a nonlinear adjustment term based on the change in fitness. The Pareto optimal solution set that satisfies all constraints is obtained by solving the improved multi-objective particle swarm optimization algorithm. Then, the compromise optimal solution is selected from the Pareto optimal solution set as the final scheduling scheme based on the fuzzy membership function.

4. The zero-carbon park energy intelligent control method based on power forecasting according to claim 3, characterized in that: The hierarchical distributed collaborative control system comprises a two-tier architecture: a park-level central controller and distributed edge computing terminals. The park-level central controller is responsible for executing the day-ahead scheduling plan and generating power allocation instructions for each sub-region based on the real-time correction plan. The distributed edge computing terminals are deployed at distributed energy devices, energy storage systems, and flexible load control nodes in each sub-region. They are responsible for collecting real-time device status data and executing autonomous control based on locally pre-set emergency control strategies when communication with the park-level central controller is interrupted. The emergency control strategies are obtained through offline training using a reinforcement learning algorithm. This reinforcement learning algorithm uses local historical operating data as training samples and aims to minimize voltage deviation and maximize renewable energy consumption to generate a state-action mapping table. In island mode, the distributed edge computing terminals look up the state-action mapping table to execute control decisions.

5. The zero-carbon park energy intelligent control method based on power forecasting according to claim 4, characterized in that: Model-control closed-loop iterative optimization is achieved by constructing a dynamic evaluation mechanism for prediction deviation, which specifically includes: calculating the deviation between the predicted and actual power values ​​at each time scale in real time and generating a prediction deviation sequence; Perform spectral analysis on the predicted deviation sequence to identify the main frequency components of the deviation and the corresponding sources of the deviation. If the source of the deviation is attributed to a sudden meteorological event, the meteorological correction module of the hybrid forecasting model will be triggered to update the parameters. If the source of the deviation is attributed to random load fluctuations, then the robust peer model of the optimization scheduling model is triggered to perform constraint relaxation adjustment. If the deviation originates from equipment response delay, then update the response compensation parameters of the hierarchical distributed collaborative control. The response compensation parameters are updated by calculating the actual response time of the most recent N control commands. With set response time Calculate the average deviation between the two, multiply it by the learning rate γ, and then combine it with the current response compensation parameter. Add them together to obtain the new response compensation parameters. .

6. The zero-carbon park energy intelligent control method based on power forecasting according to claim 5, characterized in that: It also includes steps for visualizing and tracing carbon flows in the park based on digital twins: constructing a digital twin model of the park's energy system, synchronously mapping multi-source data, power forecasting results, optimized scheduling schemes, and real-time control data to the digital twin model; in the digital twin model, constructing a carbon flow topology network of the park based on graph computing algorithms, tracking the source and destination of carbon flows at each node, and calculating the real-time carbon footprint of each energy-consuming unit; the real-time carbon footprint is calculated as follows: for any energy-consuming unit, from the initial time to the current time t, its power consumption is... Real-time carbon potential of access nodes Integrating the product of the two values ​​yields the cumulative carbon footprint of the energy-consuming unit. ; When the carbon potential of a node exceeds the warning threshold, the key path and main contributing equipment for carbon emission exceedance are located by reverse tracing through the carbon flow topology network, and carbon emission reduction auxiliary decision-making suggestions are generated.

7. A zero-carbon industrial park energy intelligent control system based on power forecasting, wherein the system is applied in the intelligent control method described in any one of claims 1-6, characterized in that: The system includes: The data acquisition module is used to acquire multi-source data from the park, including historical power load data, weather forecast data, renewable energy power generation data, and carbon emission monitoring data. The hybrid forecasting module is used to perform multi-time-scale power forecasting based on multi-source data and through a pre-built hybrid forecasting model, generating ultra-short-term power forecasting sequences, short-term power forecasting sequences, and medium- and long-term power forecasting sequences. The optimized scheduling model construction module is used to construct an optimized scheduling model for the park's energy system that takes into account carbon flow coupling characteristics, based on ultra-short-term power forecast sequences, short-term power forecast sequences, and medium- and long-term power forecast sequences. The solution module is used to solve the optimization scheduling model and generate day-ahead scheduling plans and real-time correction plans that include multi-terminal coordination of source, grid, load and storage. The hierarchical distributed control module is used to perform hierarchical distributed collaborative control of distributed energy equipment, energy storage systems and flexible loads in the park through an edge computing gateway based on the day-ahead scheduling plan and the real-time correction plan, and to collect control response data in real time. The closed-loop optimization module is used to dynamically adjust the parameters of the hybrid prediction model and the constraints of the optimized scheduling model based on the deviation between the control response data and the corresponding power forecast value, forming a model-control closed-loop iterative optimization.

Images