Optimization method and system based on large power dispatch model

Abstract

The present application discloses a tuning method and system based on a large model of power dispatch, which belongs to the field of power dispatch technology. The method includes the following steps: obtaining historical power dispatch knowledge, dividing the historical power dispatch knowledge into regional knowledge and global knowledge based on regional characteristics; constructing a first teacher model based on the differences between global knowledge and regional knowledge, and constructing a student model based on the minimum fluctuation cost; obtaining regional dispatch relevance based on historical power dispatch knowledge, and constructing a second teacher model based on regional dispatch relevance; associating the first teacher model, the second teacher model and the student model based on knowledge distillation to construct a large model of power dispatch; outputting a tuning strategy based on the large model of power dispatch and current power data. Beneficial effects of the present application: The output of the power tuning strategy takes into account both pertinence and efficiency.

Description

Technical Field

[0001] The present application relates to the technical field of electric power dispatching, and in particular to an optimization method and system based on a large electric power dispatching model. Background Art

[0002] In the operation and management of power systems, power dispatch is crucial to ensuring the stability, reliability and economy of power supply. However, traditional power dispatch models are usually designed based on certain load forecasts and power supply structures, and have poor adaptability to system changes. When clean energy such as wind power, photovoltaic power and other volatile power sources are connected in large quantities and flexible loads participate in power dispatch, there are significant defects in dispatch flexibility and optimized dispatch accuracy.

[0003] In related technologies, a unified dispatching model is usually built based on the power dispatching data of all regions. However, due to the limitations of neural network model training, some special power dispatching data will be ignored or covered during training, resulting in the unique dispatching mode of the region being ignored, resulting in reduced targeting. Building a separate power dispatching model for each region will cause too much global dispatching data to be repeatedly trained by each regional model, resulting in too much redundant calculations. Each region needs to undertake complex model training, which is inefficient. At the same time, the differences in the descriptions of each region during the global tuning process will also make it difficult to match the global tuning, and the professionalism of the dispatchers is required to be high.

[0004] The patent "Multi-regional power system economic environment dispatching method and system based on knowledge distillation", publication number: CN116306310A, publication date: June 23, 2023, specifically discloses the method including: obtaining real-time operation data and historical dispatching data of multi-regional power grids; constructing a multi-objective optimization model with minimizing fuel costs and minimizing pollution emissions as objective functions; constructing an offline support vector regression agent model; and then using the freeze-fine-tuning knowledge distillation method to construct an online agent model; using the improved fast non-dominated genetic algorithm-III to iteratively solve non-dominated solutions to obtain the Pareto optimal frontier containing all non-dominated solutions; using the membership function to calculate the membership values ​​of all non-dominated solutions to obtain the global optimal solution, and using the global optimal solution as the decision-making basis for the joint optimization dispatching of the economic environment of multi-regional power systems in the current period. Although this scheme uses the freeze-fine-tuning knowledge distillation method to construct an online agent model to reduce the complexity of the solution, the scheme still uses global data to optimize and calculate the dispatching strategies in multiple regions, and has poor pertinence.

[0005] The patent "Power Dispatching Method and Dispatching System Based on Knowledge Distillation and Federated Learning", publication number: CN117498358A, publication date: February 2, 2024, specifically discloses that by constructing corresponding intelligent dispatching models for each regional platform's historical electricity consumption data, using each intelligent dispatching model as a large model to train a small model of the aggregate model, and then obtaining the model parameters of the aggregate model, the model parameters of the aggregate model are sent to each intelligent dispatching model, and then the historical electricity consumption data of each regional platform can be used to update and iterate each intelligent dispatching model without the intelligent dispatching models interacting with each other, and finally the power consumption of the current region is predicted based on the intelligent dispatching model of the current regional platform. Although this solution constructs an intelligent dispatching model for each regional platform to ensure the targeted dispatching, each regional platform needs to undertake a complex training process, which is inefficient. Summary of the invention

[0006] In view of the fact that power tuning in the prior art cannot take into account both the tuning targeting and the tuning efficiency, the present application provides a tuning method and system based on a large power dispatching model. By constructing a first teacher model according to the differences between global knowledge and regional knowledge, constructing a second teacher model according to the regional dispatching relevance, and constructing a student model according to the minimum fluctuation cost, it is ensured that the large power dispatching model not only learns the global power dispatching knowledge and relevance, but also retains the regional individuality, so that the dispatching is more in line with the actual regional situation while satisfying the basic dispatching knowledge. In addition, the knowledge distillation technology is used to train the global knowledge matching and regional dispatching relevance calculation through the teacher model, and the student model only needs to undertake the calculation of the regional characteristics, thereby reducing the calculation complexity of regional dispatching and improving the dispatching efficiency.

[0007] To achieve the above-mentioned technical objectives, a technical solution provided in the present application is a tuning method based on a large power dispatching model, comprising the following steps: obtaining historical power dispatching knowledge, and dividing the historical power dispatching knowledge into regional knowledge and global knowledge based on regional characteristics; constructing a first teacher model based on the differences between global knowledge and regional knowledge, and constructing a student model based on the minimum fluctuation cost; obtaining regional dispatching relevance based on historical power dispatching knowledge, and constructing a second teacher model based on the regional dispatching relevance; associating the first teacher model, the second teacher model and the student model based on knowledge distillation to construct a large power dispatching model; and outputting a tuning strategy based on the large power dispatching model and current power data.

[0008] Furthermore, the method of constructing a first teacher model based on the differences between global knowledge and regional knowledge and constructing a student model based on the minimum fluctuation cost includes: constructing the first teacher model by outputting the semantic matching probability of global knowledge and regional knowledge through semantic feature recognition and similarity feature matching; constructing a tuning target optimization function with the minimum fluctuation cost, and constructing a student model based on the tuning target optimization function.

[0009] Furthermore, the method of acquiring regional scheduling relevance based on historical power scheduling knowledge and constructing a second teacher model based on regional scheduling relevance includes: acquiring a first regional scheduling relevance based on historical power scheduling influence relationship, and constructing a second regional scheduling relevance based on the semantic similarity of regional knowledge; and constructing a second teacher model with the first regional scheduling relevance and the second regional scheduling relevance.

[0010] Furthermore, the construction of the second teacher model based on the first correlation of regional scheduling and the second correlation of regional scheduling includes: outputting the position correlation probability based on the first correlation of regional scheduling; outputting the semantic correlation probability based on the second correlation of regional scheduling; outputting the region correlation probability based on the position correlation probability and the semantic correlation probability to construct the second teacher model.

[0011] Furthermore, constructing a tuning target optimization function with the lowest fluctuation cost and constructing a student model based on the tuning target optimization function includes: constructing a tuning target optimization function with equipment change ratio, regional impact ratio and deviation probability, and constructing a student model with the tuning target optimization function.

[0012] Furthermore, the method of associating the first teacher model, the second teacher model and the student model based on knowledge distillation to construct a large power dispatching model also includes: constructing a global dynamic adjustment layer based on association judgment; and constructing a large power dispatching model based on the global dynamic adjustment layer and the student model.

[0013] Furthermore, the output tuning strategy based on the power dispatching large model and current power data includes: outputting a regional tuning strategy based on the current power data and each student model; outputting an associated area based on the regional tuning strategy, and outputting a global tuning strategy based on the regional tuning strategy of the associated area.

[0014] Furthermore, outputting associated areas according to the regional tuning strategy and outputting the global tuning strategy according to the regional tuning strategy of the associated areas includes: outputting associated areas according to the regional tuning strategy; if there is a unique regional tuning strategy for the associated areas, obtaining the global tuning strategy based on the unique regional tuning strategy; if there are multiple regional tuning strategies for the associated areas, outputting the global tuning strategy based on the multiple regional tuning strategies and tuning weights.

[0015] Furthermore, it also includes: obtaining weighted fine-tuning timing based on timing operation characteristics according to historical power dispatching data; fine-tuning the timing according to the weights and optimizing the weights to fine-tune the power dispatching model.

[0016] Furthermore, the method of obtaining weight fine-tuning timing based on timing operation characteristics according to historical power dispatching data includes: obtaining equipment dispatching cycles related to timing in historical power dispatching data according to equipment operation repetitions, and obtaining the timing operation characteristics of each device in the equipment dispatching cycle according to the equipment dispatching cycle; obtaining the current operating timing, and using the timing operation characteristics of each device to obtain weight fine-tuning timing for each device that matches the current operating timing.

[0017] Furthermore, the method of utilizing the timing operation characteristics of each device to obtain the weight fine-tuning timing of each device that matches the current operation timing also includes: obtaining the scheduling compensation period of each device that matches the current timing according to the timing operation characteristics of each device, and outputting the weight fine-tuning timing and weight fine-tuning value according to the compensation node of the scheduling compensation period of each device.

[0018] Another technical solution provided by the present application is a tuning system based on a large model of power dispatching, which is used to implement the method as described above, including: a data acquisition unit, used to acquire power dispatching knowledge and current power data; a data modeling unit, including a first modeling unit for constructing a first teacher model, a second modeling unit for constructing a second teacher model, and a third modeling unit for constructing a student model, for constructing a large model of power dispatching; an analysis and tuning unit, used to output a tuning strategy based on the large model of power dispatching and current power data.

[0019] The beneficial effects of the present application are as follows: 1. Based on the first teacher model and the second teacher model, training with global knowledge content such as the difference between global description and regional description, the difference between global tuning and regional tuning, and the scheduling impact between regions, reducing the student model training data, replacing the repeated training of the student model with the unified training of the teacher model, reducing the redundant training of the student model, simplifying the training process, improving efficiency, and retaining regional characteristics with the targeted optimization of the student model.

[0020] 2. The fluctuation cost of the tuning process is calculated by the proportion of equipment changes in the region, the proportion of impact on other regions, and the probability of tuning deviation in the tuning area during the regional tuning process, so as to achieve stable adjustment of the region while minimizing the proportion of equipment changes and reducing the impact on other regions.

[0021] 3. Output the associated regions according to the regional tuning strategy, and output the global tuning strategy according to the regional tuning strategy of the associated regions. When performing global tuning, only the associated regions need to be adjusted, and the unassociated regions do not need to participate in the global tuning, thereby improving the tuning adaptability and the tuning efficiency.

[0022] 4. Use the timing operation characteristics of the power system to obtain the weight update timing of the model tuning weights. On the basis of improving the adaptability and accuracy of the model, avoid adjusting the model-irrelevant weights, reduce the complexity of model training, and improve the efficiency of model fine-tuning.

[0023] 5. Fine-tune the timing and value of the weight output of the compensation node according to the compensation period of each device scheduling to ensure a smooth transition of the equipment during the tuning process and ensure the stability of the power grid.

[0024] 6. While retaining the semantic description of regional characteristics during the tuning process, global macro-adjustment can be achieved, the accuracy of global tuning can be improved, the adaptability of regional tuning can be ensured, the requirements for dispatchers can be reduced, and tuning differences caused by discrepancies in terminology can be avoided. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 This is a flow chart of the optimization method based on the large power dispatching model of this application.

[0026] Figure 2 This is a structural diagram of an embodiment of the first teacher basic model of this application. DETAILED DESCRIPTION

[0027] In order to make the purpose, technical solutions and advantages of the present application more clear, the present application is further described in detail below in conjunction with the drawings and examples. It should be understood that the specific implementation method described here is only an optimal embodiment of the present application, which is only used to explain the present application and does not limit the scope of protection of the present application. All other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present application.

[0028] like Figure 1 As shown, as the first embodiment of the present application, the tuning method based on the large power dispatching model includes the following steps:

[0029] Acquire historical power dispatching knowledge, and divide it into regional knowledge and global knowledge based on regional characteristics;

[0030] The first teacher model is constructed based on the differences of global knowledge and regional knowledge, and the student model is constructed based on the minimum fluctuation cost;

[0031] Acquire regional dispatch relevance based on historical power dispatch knowledge, and build a second teacher model based on regional dispatch relevance;

[0032] Based on knowledge distillation, the first teacher model, the second teacher model and the student model are associated to build a large power dispatch model;

[0033] Output power optimization strategy based on the power dispatching model and current power data.

[0034] In this embodiment, the first teacher model is constructed based on the differences in global knowledge and regional knowledge, the second teacher model is constructed based on the regional scheduling relevance, and the student model is constructed based on the minimum fluctuation cost, thereby ensuring that the large power scheduling model not only learns the global power scheduling knowledge and relevance, but also retains the regional individuality, so that the scheduling is more in line with the actual regional situation while satisfying the basic scheduling knowledge, and uses knowledge distillation technology to reduce the computational complexity of regional scheduling and improve scheduling efficiency.

[0035] The historical power dispatching knowledge at least includes historical power operation and dispatching data, power dispatching and operation procedures, power equipment parameters, historical fault cases, and special working condition processing strategies. The part of the historical power dispatching knowledge that highlights regional characteristics is divided into the regional knowledge of the region, and the part of the historical power dispatching knowledge that highlights global characteristics is divided into global knowledge.

[0036] Knowledge distillation is a model compression and acceleration technology in the field of machine learning. It transfers the "knowledge" learned by a complex, large and high-performance model (usually called the teacher model) to a relatively simple and small model (student model), so that the student model can imitate the behavior and output of the teacher model as much as possible, thereby reducing the model complexity and computing resource consumption while maintaining certain performance. In this embodiment, the first teacher model is constructed based on the differences in global knowledge and regional knowledge, and the student model is constructed based on the minimum fluctuation cost, including:

[0037] The first teacher model is constructed by outputting the semantic matching probability of global knowledge and regional knowledge through semantic feature recognition and similarity feature matching;

[0038] A tuning target optimization function is constructed with the lowest volatility cost, and a student model is constructed based on the tuning target optimization function.

[0039] The semantic matching probability of regional knowledge and global knowledge is obtained through the semantic feature matching similarity between global knowledge and regional knowledge, so that global macro adjustment can be achieved while retaining the semantic description of regional characteristics during the tuning process, improving the accuracy of global tuning, ensuring the adaptability of regional tuning, reducing the requirements for dispatchers, and avoiding tuning differences caused by discrepancies in terminology descriptions. At the same time, the tuning target optimization function is constructed with the lowest fluctuation cost, and the student model corresponding to each region is constructed to achieve targeted scheduling optimization for each region.

[0040] like Figure 2As shown, the first teacher model is constructed by outputting the semantic matching probability of global knowledge and regional knowledge through semantic feature recognition and similarity feature matching, including:

[0041] Construct the global feature recognition layer of the first teacher basic model based on the Transformer architecture and global knowledge;

[0042] Construct the regional feature recognition layer of the first teacher base model based on the Transformer architecture and regional knowledge;

[0043] Constructing the similarity matching layer of the first teacher basic model based on the correlation between global knowledge and regional knowledge;

[0044] The global feature recognition layer, regional feature recognition layer and similarity matching layer of the first teacher basic model are trained based on the global knowledge and regional knowledge to obtain the first teacher model.

[0045] Constructing a tuning target optimization function with the lowest volatility cost, and constructing a student model based on the tuning target optimization function includes:

[0046] The tuning target optimization function is constructed based on the equipment change ratio, regional impact ratio and deviation probability, and the student model is constructed based on the tuning target optimization function.

[0047] In this embodiment, the first teacher model is used to learn the semantic expression of global knowledge and regional knowledge, and obtain the similarity between the expression of global knowledge and regional knowledge, such as using cosine similarity, dot product similarity, etc. to calculate the similarity between the global semantic feature vector and the regional semantic feature vector, and use the similarity as the semantic matching probability, so as to highlight the correlation between the regional description and the global description, and avoid the influence of regional language style and expression habits on global tuning. At the same time, the minimum fluctuation cost is calculated using the equipment change ratio, regional impact ratio and deviation probability, so as to construct the tuning target optimization function and obtain the tuning model considering the minimum fluctuation cost. It can be understood that in this embodiment, the student model does not participate in the calculation of power grid load and power grid output, and can be directly obtained according to the existing power grid load prediction model and power grid output prediction model. The student model outputs the tuning strategy according to the power grid load data and power grid output data combined with the tuning target optimization function model. Of course, in other embodiments, the student model and the power grid load prediction model and the power grid output prediction model can be integrated to obtain a fusion model to improve the integration of the model.

[0048] The optimization function of the tuning objective is constructed based on the equipment change ratio, regional impact ratio and deviation probability:

[0049] ;

[0050] ;

[0051] ;

[0052] ;

[0053] ;

[0054] ;

[0055] in, represents the tuning target optimization function, Indicates the equipment change ratio, represents the regional impact ratio, represents the probability of deviation; represents the number of power equipment that needs to be changed for optimization in area i, represents the number of all power equipment in area i; Representation and area The number of co-tuned regions j, Indicates the total number of regions; represents the probability of load forecast deviation, represents the probability of output prediction deviation, represents the tuning deviation probability of region i, represents the tuning deviation probability of region j, Indicates the total number of historical tuning samples for region i, represents the mth historical tuning sample of region i, represents the actual value of the mth historical tuning sample in region i, represents the tuning output value of the mth historical tuning sample in region i, is the preset deviation threshold of region i, represents the total number of historical tuning samples in region j, represents the mth historical tuning sample of region j, represents the actual value of the mth historical tuning sample in region j, represents the tuning output value of the mth historical tuning sample in region j, is the preset deviation threshold of region j, Represents an indicator function that returns 1 if the condition is true, otherwise it returns 0.

[0056] In this embodiment, the fluctuation cost of the tuning process is calculated by the proportion of equipment changes in the region, the proportion of impact on other regions, and the probability of tuning deviation in the tuning region during the regional tuning process, so as to achieve stable adjustment of the region while minimizing the proportion of equipment changes and reducing the impact on other regions.

[0057] In some cases, , Both are set to 0.

[0058] At the same time, the regional dispatch relevance is obtained based on the historical power dispatch knowledge, and the second teacher model is constructed based on the regional dispatch relevance, including:

[0059] The first relevance of regional dispatch is obtained according to the historical power dispatch impact relationship, and the second relevance of regional dispatch is constructed according to the semantic similarity of regional knowledge;

[0060] The second teacher model is constructed with the first correlation of regional scheduling and the second correlation of regional scheduling.

[0061] In this embodiment, historical power dispatching knowledge is used to obtain the historical power dispatching influence relationship. For example, when a certain device in a certain area is adjusted, a certain device in another area will be adjusted synchronously. It is considered that there is an influence relationship between the two areas, and the first correlation of regional dispatching between areas is constructed according to the proportion of regional synchronous dispatching. At the same time, the second correlation of regional dispatching is constructed according to the semantic similarity between each area, and the second correlation between each area is obtained through the correlation between language and regional habits, so as to highlight the differences in behavioral habits between adjacent regions. For example, there are three areas A, B, and C. Area A is adjacent to Area B, and Area B is adjacent to Area C. The semantic similarity between Area A and Area B is lower than the semantic similarity between Area B and Area C. It is considered that Area B and Area C have a higher correlation.

[0062] The second teacher model of the influence probability between regions is constructed by the first correlation of regional scheduling and the second correlation of regional scheduling to reflect the regional influence relationship and the correlation of behavioral habits between regions, that is, the second teacher model is constructed by the first correlation of regional scheduling and the second correlation of regional scheduling, including:

[0063] Outputting the location-related probability with the first correlation of regional scheduling;

[0064] Outputting semantic relevance probability with regional scheduling second relevance;

[0065] The region-related probability is output based on the position-related probability and the semantic-related probability to construct the second teacher model.

[0066] Based on the first teacher model and the second teacher model, training with global knowledge content such as the difference between global description and regional description, the difference between global tuning and regional tuning, and the scheduling impact between regions reduces the student model training data, replaces the repeated training of the student model with the unified training of the teacher model, reduces the redundant training of the student model, simplifies the training process, improves efficiency, and retains regional characteristics through targeted optimization of the student model.

[0067] Furthermore, based on knowledge distillation, the first teacher model, the second teacher model and the student model are associated to construct a large power dispatch model including:

[0068] The semantic matching probability of the first teacher model and the regional correlation probability of the second teacher model are used as soft labels of the student model;

[0069] Construct a global dynamic adjustment layer based on association judgment;

[0070] A large power dispatching model is constructed based on the global dynamic adjustment layer and the student model.

[0071] By using semantic matching probability and regional correlation probability as soft labels for the student model, it can learn the professional expression of scheduling and the impact of scheduling on other areas. When scheduling in this area, it can simultaneously output regional characteristic descriptions and global professional descriptions, making it easier for dispatchers to understand the output scheduling strategy and unify scheduling information. At the same time, based on the location correlation and behavioral habit correlation between regions, the impact of the scheduling process on other regions is obtained, further improving the accuracy of the impact judgment on similar areas and improving the adaptability of the scheduling strategy.

[0072] Among them, the optimization strategies based on the power dispatching model and the current power data output include:

[0073] Output regional optimization strategies based on current power data and various student models;

[0074] The associated regions are output according to the regional tuning strategy, and the global tuning strategy is output according to the regional tuning strategy of the associated regions.

[0075] In this embodiment, the student model learns the regional correlation probability of the second teacher model, so that when the regional tuning strategy is output, the location correlation and behavioral habit correlation of the remaining regions are included in the calculation of the fluctuation cost. When the behavioral habits in adjacent regions are similar, the load usage will also have the same trend, such as whether to use heating at a certain temperature, etc. The correlation of behavioral habits in adjacent regions is highlighted through regional description characteristics, so that in adjacent regions with high correlation, when power dispatch occurs in one region, the power dispatch strategy of another region is output synchronously, which is convenient for dispatchers to refer to. Or, when there is a conflict in the tuning strategies of two regions, global tuning is performed, such as the additional inclusion of clean energy.

[0076] Outputting the associated area based on the regional tuning strategy and outputting the global tuning strategy based on the regional tuning strategy of the associated area include:

[0077] Output the associated regions according to the regional optimization strategy;

[0078] If the associated area has a unique regional tuning policy, the global tuning policy is obtained based on the unique regional tuning policy;

[0079] If there are multiple regional tuning strategies for the associated region, the global tuning strategy is output based on the multiple regional tuning strategies and tuning weights.

[0080] When there are multiple regional tuning strategies in an associated area, unified tuning is performed according to the tuning weights. The tuning weights can be set according to the priority of the power equipment, such as the access priority of clean energy is higher than the priority of increasing the output power of traditional output equipment, or the priority of flexible loads is lower than that of rigid loads. Since only the associated areas need to be adjusted during global tuning, and unassociated areas do not need to participate in global tuning, the tuning efficiency is improved while improving the tuning adaptability. It is understandable that since each learning model has learned the matching probability of the global knowledge semantic features and the regional knowledge semantic features, the regional tuning strategy output by the learning model includes a regional characteristic description and a global professional description, so that global tuning is performed based on the global professional description in the process of joint tuning of multiple regional tuning strategies, avoiding tuning deviations caused by description discrepancies.

[0081] As a second embodiment of the present application, the tuning method based on the large power dispatch model also includes:

[0082] Obtain weights and fine-tune the timing based on the timing operation characteristics according to historical power dispatch data;

[0083] Fine-tune the timing according to the weights and fine-tune the power dispatching model by optimizing the weights.

[0084] The timing operation characteristics of the power system are used to obtain the weight update timing of the model tuning weights. On the basis of improving the adaptability and accuracy of the model, the adjustment of model-irrelevant weights is avoided, the complexity of model training is reduced, and the efficiency of model fine-tuning is improved.

[0085] Specifically, obtaining the weight fine-tuning timing based on the timing operation characteristics according to the historical power dispatching data includes:

[0086] Obtain the timing operation characteristics of each device based on the device dispatch cycle in the historical power dispatch data;

[0087] The weighted fine-tuning timing is calculated based on the timing operating characteristics of each device.

[0088] In this embodiment, the weight fine-tuning timing of different tuning weights is set according to the timing operation characteristics of each device. For example, when the operation of the photovoltaic output equipment is unstable, the tuning weight of the photovoltaic output equipment is reduced; when the operation of the photovoltaic equipment is stable, the tuning weight of the photovoltaic output equipment is increased; and the dynamic tuning weight of each device is obtained according to the participation of the device at different time periods in the tuning stage.

[0089] In other embodiments, the tuning weight is obtained according to the equipment participation, equipment failure rate, equipment maintenance cost and equipment substitutability. The equipment participation is calculated according to the participation ratio of the equipment in the historical tuning stage; the equipment failure rate is calculated according to the failure data ratio of the equipment in the historical operation data; the equipment maintenance cost includes the maintenance time cost and the maintenance economic cost, which is calculated according to the time required for equipment maintenance and the required cost. The equipment substitutability is the degree to which the equipment can be replaced in the entire power system. It is calculated according to the similarity of the equipment function and the equipment power to obtain the equipment substitutability, that is, the number of equipment with similar functions and powers to the equipment that can replace its work when the equipment fails to work properly is used as the substitutability of the equipment. The equipment participation, equipment failure rate, equipment maintenance cost and equipment substitutability are normalized and calculated based on the exponential function to obtain the tuning weight.

[0090] Specifically, the device participation is:

[0091] ;

[0092] in, is the device engagement of device n, is the total number of times device n participates in tuning, The total number of historical tunings.

[0093] The equipment failure rate is:

[0094] ;

[0095] in, is the device failure rate of device n, is the number of failures of device n in historical operation data, is the total number of times device n runs in the historical running data.

[0096] Equipment maintenance costs are:

[0097] ;

[0098] in, is the equipment maintenance cost of equipment n, is the conversion factor, The time required for maintenance of device n, Costs required for equipment maintenance.

[0099] Equipment alternatives are:

[0100] ;

[0101] in, is the device substitutability of device n, It is the number of devices with similar functions and power as device n and that can replace it when it fails to work properly.

[0102] The tuning weights are obtained based on the equipment participation, equipment failure rate, equipment maintenance cost, and equipment substitutability:

[0103] ;

[0104] ;

[0105] ;

[0106] ;

[0107] ;

[0108] ;

[0109] in, is the normalized value of device engagement of device n, is the minimum value of device participation, is the maximum device participation, is the weight coefficient of device participation; is the normalized value of the device failure rate of device n, is the minimum value of equipment failure rate, is the maximum value of equipment failure rate, is the weight coefficient of equipment failure rate; is the normalized value of equipment maintenance cost of equipment n, is the minimum equipment maintenance cost, is the maximum equipment maintenance cost, is the weight coefficient of equipment maintenance cost; For equipment The normalized value of device engagement, is the minimum value of device participation, is the maximum device participation, is the weight coefficient of device participation.

[0110] It is understandable that during the tuning process, a second normalization calculation is performed based on the tuning weights of each device, that is, the tuning weight of each device is divided by the sum of all tuning weights to obtain the final tuning weight.

[0111] The timing operation characteristics of each device are obtained based on the device dispatch cycle in the historical power dispatch data, including:

[0112] The equipment dispatch cycle related to the timing in the historical power dispatch data is obtained according to the equipment operation repetition situation, and the timing operation characteristics of each equipment in the equipment dispatch cycle are obtained according to the equipment dispatch cycle.

[0113] The weight fine-tuning timing that calculates the tuning weight based on the timing operation characteristics of each device includes:

[0114] The current running timing is obtained, and the weighted fine-tuning timing of each device that matches the current running timing is obtained by using the timing running characteristics of each device.

[0115] In this embodiment, the change of equipment operation during a equipment scheduling cycle is obtained through the repetitive situation of equipment operation. For example, photovoltaic output, flexible load and other fluctuating power equipment will still have operation changes after scheduling. The difference between photovoltaic output capacity during the day and at night will inevitably cause the existence of night compensation equipment in the scheduling process, such as energy storage equipment connected at night or wind equipment connected at night to compensate for the difference in photovoltaic output. Therefore, the photovoltaic equipment has a timing operation characteristic affected by light. Under different timings, the output time period of the photovoltaic equipment is not the same. Therefore, the timing of the fluctuation of the photovoltaic equipment output is used as the weight fine-tuning timing. Under this timing, the impact of the decline in the output of the photovoltaic equipment needs to be considered. Similarly, for flexible loads, the entire cycle of flexible load connection and non-connection is used as the timing operation characteristic of the flexible load equipment, and the time nodes of flexible load connection and non-connection are used as the weight fine-tuning timing.

[0116] Fine-tuning the timing according to the weight and optimizing the weight to fine-tune the large model of power dispatch include:

[0117] Calculate the tuning weight under the weight fine-tuning timing based on the historical timing corresponding to the weight fine-tuning timing;

[0118] The tuning weights of the power dispatching large model are updated according to the tuning weights under the current timing and the tuning weights under the weight fine-tuning timing.

[0119] Considering the dynamically changing equipment, if only the tuning weights under the current timing are calculated, that is, the equipment participation, equipment failure rate, equipment maintenance cost and equipment substitutability under the current timing, the tuning results will not be optimal when the equipment changes. Therefore, the tuning weights when the equipment changes dynamically, that is, when the weight is fine-tuned, are calculated at the same time, and a comprehensive calculation is performed as the tuning weight of the large power dispatch model to ensure that the tuning result is still the optimal solution under fluctuations. For example, for photovoltaic equipment, in a dark environment, the output needs to be replaced by energy storage equipment. At this time, the timing of switching / complementary output between photovoltaic equipment and energy storage equipment is the weight fine-tuning timing. At this time, it is necessary to calculate the equipment participation, equipment failure rate, equipment maintenance cost and equipment substitutability of the energy storage equipment to calculate the tuning weight under the weight fine-tuning timing to improve the tuning accuracy.

[0120] The tuning weights of the power dispatching model are updated according to the tuning weights under the current time sequence and the tuning weights under the weight fine-tuning time sequence:

[0121] ;

[0122] in, is the tuning weight of the updated power dispatching model, The duration of the current sequence operation. is the tuning weight under the current timing, To fine-tune the weight, the duration of the timing device operation is Tune weights for weight fine-tuning timing.

[0123] It can be understood that this embodiment takes the case where the equipment has one fluctuation as an example. In other cases, if the equipment has multiple fluctuation adjustments before the repeated operation state occurs, there will be multiple weight fine-tuning sequences and tuning weights under the weight fine-tuning sequences. The tuning weights of the power dispatching model are updated according to the proportion of the continuous operation time of the equipment after multiple weight fine-tuning sequences.

[0124] In other cases, considering the impact on system stability, using the timing operation characteristics of each device to obtain the weight fine-tuning timing of each device that matches the current operation timing also includes:

[0125] According to the timing operation characteristics of each device, the scheduling compensation period of each device matching the current timing is obtained, and the weight fine-tuning timing and weight fine-tuning value are output according to the compensation node of each device scheduling compensation period.

[0126] In this embodiment, the timing operation characteristics of each device are the changes in the output or input of the device in a device scheduling cycle. Since the sudden connection or sudden connection of the device may affect the stability of the system, in some cases, the device will be connected step by step to complete the power scheduling, and the grid tuning will be achieved while ensuring the stability of the grid. The compensation period for each device scheduling is the period when the output or input of the device continues to change during the historical tuning process. Each change node is used as a compensation node, and the weight is fine-tuned, and the weight fine-tuning value is output according to the degree of change. For example, if a certain device has been running in the previous period, but its output needs to be reduced and the output of another device needs to be increased in the regional tuning strategy, at this time, the tuning weight of the device is reduced according to the weight fine-tuning timing and the weight fine-tuning value, and the tuning weight of another device is increased, so as to try to ensure a smooth transition of the device during the tuning process and ensure the stability of the grid.

[0127] Specifically, the compensation node output weight fine-tuning timing and weight fine-tuning value according to the compensation period of each device scheduling include:

[0128] Fine-tune the timing according to the compensation node output weights of each device scheduling compensation period;

[0129] Obtain the weight fine-tuning value of each weight fine-tuning timing according to the impact of device fluctuation timing and the tuning weight.

[0130] The impact of equipment fluctuation sequence includes voltage fluctuation, power flow change, system frequency response, etc. The power system analysis tools such as power flow calculation and transient stability analysis are used to evaluate the impact of equipment connection and disconnection on the stability of the power grid at different sequences. For each weight fine-tuning sequence, the impact of fluctuations during equipment connection and disconnection on the stability of the power grid is calculated, and the weight fine-tuning values ​​under each weight fine-tuning sequence that satisfies the final adjustment to the target tuning weight and the minimum total impact on the power grid stability are calculated using optimization algorithms (such as genetic algorithms, particle swarm optimization, etc.), so as to minimize the impact on the stability of the power grid while meeting the tuning requirements.

[0131] At this point, build a weight target optimization function based on the impact of device fluctuation timing and tuning weights:

[0132] ;

[0133] The construction constraints are:

[0134] ;

[0135] in, is the weight objective optimization function, is the weight fine-tuning value at time t, is the voltage fluctuation influence coefficient at time t, is the influence coefficient of power flow change at time t, is the influence coefficient of system frequency response at time t, and T is the total tuning time.

[0136] In some other embodiments, constructing the first teacher model based on the differences between global knowledge and regional knowledge further includes:

[0137] The global tuning strategy and regional tuning strategy based on the same time series output the regional tuning fluctuation probability;

[0138] The first teacher model is constructed with regional tuning fluctuation probability and semantic matching probability.

[0139] According to the difference between the regional tuning strategy and the global tuning strategy, the regional tuning fluctuation probability corresponding to the region is obtained, so as to obtain the region that deviates more from the global tuning strategy. Therefore, when there are multiple regional tuning strategies, the global tuning strategy is adjusted according to the regional tuning fluctuation probability, weight fine-tuning timing, and weight fine-tuning value, so that the regional tuning strategy of the region with greater difference from the global tuning in the historical tuning process accounts for a smaller proportion, and the regional tuning strategy of the region with smaller difference accounts for a larger proportion, and the global tuning strategy is output in combination with the tuning weight of the device. Since the regional tuning strategy output by each student model is targeted at the characteristics of its own region, even if it only accounts for a small proportion in the global tuning strategy, it is more targeted than the common tuning of all regions, and will not be affected by the tuning of many irrelevant regions. At the same time, it reduces the calculation amount of global tuning and improves the efficiency of global tuning.

[0140] As a third embodiment of the present application, a tuning system based on a large power dispatching model includes:

[0141] Data collection department, used to obtain power dispatch knowledge and current power data;

[0142] A data modeling unit, comprising a first modeling unit for constructing a first teacher model, a second modeling unit for constructing a second teacher model, and a third modeling unit for constructing a student model, for constructing a large power dispatching model;

[0143] The analysis and tuning department is used to output tuning strategies based on the power dispatching model and current power data.

[0144] In this embodiment, the knowledge distillation technology is used to undertake the complex semantic recognition and scheduling correlation acquisition training process by the first teacher model and the second teacher model. The student model obtains the semantic matching probability and regional correlation probability output by the first teacher model and the second teacher model to iteratively calculate the tuning target optimization function, and obtains the tuning strategy that meets the tuning target optimization function. It simultaneously outputs the description that conforms to the regional characteristics and the description of the global profession to facilitate the understanding and preservation of the tuning strategy. It ensures that the tuning strategy conforms to the influence relationship between regions without the student model having to perform too much complex calculations, thereby achieving the purpose of global tuning and improving the efficiency of tuning.

[0145] In some embodiments, the data modeling unit is connected to the load forecasting model and the output forecasting model, and the data output by the load forecasting model and the output forecasting model are used as the current power data, so that the student model outputs the optimization strategy according to the load forecast value and the output forecast value. In other embodiments, the second modeling unit is responsible for the construction of the load forecasting model and the output forecasting model, and the regional correlation probability output by the second teacher model is used as a reference for the regional correlation impact in the load forecasting model and the output forecasting model, thereby reducing repeated calculation steps.

[0146] As a fourth embodiment of the present application, a computer-readable storage medium is used to store a computer program or instruction. When the computer program or instruction is executed by a processing device, the above-mentioned tuning method based on the large power dispatch model is implemented. The computer-readable storage medium can be any available medium that can be stored by a computing device or a data storage device such as a data center containing one or more available media. The available medium can be a magnetic medium (for example, a floppy disk, a hard disk, a tape), an optical medium (for example, a DVD), or a semiconductor medium (for example, a solid-state hard disk), etc.

[0147] The specific implementation method described above is a preferred implementation method of the tuning method and system based on the large-scale power dispatching model of the present application, and is not intended to limit the specific implementation scope of the present application. The scope of the present application includes but is not limited to the specific implementation method. All equivalent changes made in accordance with the shape and structure of the present application are within the protection scope of the present application.

Claims

1. The optimization method based on the large power dispatching model is characterized by: The steps include: Acquire historical power dispatching knowledge, and divide it into regional knowledge and global knowledge based on regional characteristics; The first teacher model is constructed based on the differences of global knowledge and regional knowledge, and the student model is constructed based on the minimum fluctuation cost; Acquire regional dispatch relevance based on historical power dispatch knowledge, and build a second teacher model based on regional dispatch relevance; Based on knowledge distillation, the first teacher model, the second teacher model and the student model are associated to build a large power dispatch model; Output tuning strategies based on the power dispatching model and current power data; The first teacher model is constructed by outputting the semantic matching probability of global knowledge and regional knowledge through semantic feature recognition and similarity feature matching; Construct a tuning target optimization function with the lowest fluctuation cost, and construct a student model based on the tuning target optimization function; The first relevance of regional dispatch is obtained according to the historical power dispatch impact relationship, and the second relevance of regional dispatch is constructed according to the semantic similarity of regional knowledge; The second teacher model is constructed with the first correlation of regional scheduling and the second correlation of regional scheduling.

2. The optimization method based on the large power dispatching model according to claim 1, characterized in that: The constructing of the second teacher model based on the first correlation of regional scheduling and the second correlation of regional scheduling includes: Outputting the location-related probability with the first correlation of regional scheduling; Outputting semantic relevance probability with regional scheduling second relevance; The region-related probability is output based on the position-related probability and the semantic-related probability to construct the second teacher model.

3. The optimization method based on the large power dispatching model according to claim 1, characterized in that: The method of constructing a tuning target optimization function with the lowest fluctuation cost and constructing a student model based on the tuning target optimization function includes: The tuning target optimization function is constructed based on the equipment change ratio, regional impact ratio and deviation probability, and the student model is constructed based on the tuning target optimization function.

4. The tuning method based on the large power dispatching model according to claim 1, characterized in that: The method of associating the first teacher model, the second teacher model and the student model based on knowledge distillation to construct a large power dispatch model further includes: Construct a global dynamic adjustment layer based on association judgment; A large power dispatching model is constructed based on the global dynamic adjustment layer and the student model.

5. The tuning method based on the large power dispatching model as claimed in claim 4, characterized in that: The optimization strategy output according to the power dispatching big model and current power data includes: Output regional optimization strategies based on current power data and various student models; The associated regions are output according to the regional tuning strategy, and the global tuning strategy is output according to the regional tuning strategy of the associated regions.

6. The optimization method based on the large power dispatching model according to claim 5, characterized in that: Outputting the associated area according to the regional tuning strategy, and outputting the global tuning strategy according to the regional tuning strategy of the associated area include: Output the associated regions according to the regional optimization strategy; If the associated area has a unique regional tuning policy, the global tuning policy is obtained based on the unique regional tuning policy; If there are multiple regional tuning strategies for the associated region, the global tuning strategy is output based on the multiple regional tuning strategies and tuning weights.

7. The optimization method based on the large power dispatching model according to claim 6, characterized in that: Also includes: Obtain weights and fine-tune the timing based on the timing operation characteristics according to historical power dispatch data; Fine-tune the timing according to the weights and fine-tune the large power dispatch model by optimizing the weights.

8. The optimization method based on the large power dispatching model according to claim 7, characterized in that: The method of obtaining the weight fine-tuning timing based on the timing operation characteristics according to the historical power dispatching data includes: According to the equipment operation repetition situation, the equipment dispatch cycle related to the timing in the historical power dispatch data is obtained, and the timing operation characteristics of each equipment in the equipment dispatch cycle are obtained according to the equipment dispatch cycle; The current running timing is obtained, and the weighted fine-tuning timing of each device that matches the current running timing is obtained by using the timing running characteristics of each device.

9. The optimization method based on the large power dispatching model according to claim 8, characterized in that: The method of using the timing operation characteristics of each device to obtain the weight fine-tuning timing of each device that matches the current operation timing also includes: According to the timing operation characteristics of each device, the scheduling compensation period of each device matching the current timing is obtained, and the weight fine-tuning timing and weight fine-tuning value are output according to the compensation node of each device scheduling compensation period.

10. A tuning system based on a large power dispatching model, used to implement the method according to any one of claims 1 to 9, characterized in that: include: Data collection department, used to obtain power dispatch knowledge and current power data; A data modeling unit, comprising a first modeling unit for constructing a first teacher model, a second modeling unit for constructing a second teacher model, and a third modeling unit for constructing a student model, for constructing a large power dispatching model; The analysis and tuning department is used to output tuning strategies based on the power dispatching model and current power data.

Images