AGC (Automatic Gain Control) regulation demand decomposition method for dynamically matching real-time operation condition of power supply

Abstract

The invention discloses an AGC regulation demand decomposition method for dynamically matching real-time operation conditions of a power supply. The AGC regulation demand decomposition method comprises the following steps: 1) constructing a power supply frequency modulation performance perception model; 2) setting a historical frequency modulation performance vector of the power supply as a standard parameter, inputting a current multi-dimensional operation condition of the power supply into the power supply frequency modulation performance perception model, and extracting an AGC instruction adaptation range of the power supply; and 3) the power supply AGC instruction adaptation range is embedded into an AGC distribution strategy, and dynamic matching between AGC instruction issuing and power supply operation conditions is realized. On the premise that the AGC distribution strategy of the embedded power supply instruction adaptation range meets the system adjustment requirement, the adjustment instruction matched with the operation condition of the power supply is distributed to each power supply, so that the response performance of the power supply instruction is guaranteed, and the frequency quality of the system is improved.

Description

Technical Field

[0001] The present invention relates to the field of secondary frequency regulation of power systems, and specifically to a method for decomposing the AGC regulation demand that dynamically matches the real-time operating conditions of power sources. Background Art

[0002] Building a new power system with new energy as the main body has become an inevitable future development trend of China's power grid. However, affected by its strong stochastic fluctuation characteristics, the large-scale grid connection of new energy will significantly increase the real-time balance burden of power sources and loads in the power grid. How to maintain the system frequency quality has also become one of the key challenges faced in the construction of the new power system.

[0003] As an important means for the power grid to maintain system frequency quality, Automatic Generation Control (AGC) forms an AGC command (i.e., the active power regulation demand of the power grid) based on the system status monitoring value, and distributes it to each frequency regulation resource in a certain manner, so as to dispatch the resource to execute the output adjustment to cope with the system power fluctuation. Among them, the AGC command distribution is an important link in determining the system frequency quality. An unreasonable distribution strategy will make it difficult for the frequency regulation resources to effectively respond to the received AGC command. On the one hand, it will lead to a decline in the power grid frequency quality. On the other hand, the resources themselves will also be evaluated due to unqualified frequency regulation performance.

[0004] The existing AGC command distribution methods mainly include heuristic-based real-time distribution methods, advanced optimization distribution methods, and phased distribution methods that combine the two, which are described as follows.

[0005] The heuristic-based real-time distribution method mainly designs rules according to a certain characteristic of the frequency regulation resource to quickly distribute the AGC command formed based on the system frequency deviation monitoring value, etc. The common rules include regulation speed ratio, regulation capacity ratio, regulation speed order, regulation cost order, etc. To consider the differential regulation characteristics of different categories of frequency regulation resources in a coordinated manner, some studies have also proposed improved methods based on AGC command decomposition, which separately send high / low frequency commands to the frequency regulation resources that match their own regulation characteristics to achieve the complementary advantages of different resources. This type of method is simple and easy to implement, so it is widely used in the industrial community. However, this method is difficult to comprehensively consider the characteristics such as regulation capacity, regulation speed, and regulation cost of the resources, and there is a lag (the resource output adjustment lags behind the power fluctuation).

[0006] The advanced optimization distribution method constructs an optimization distribution model that comprehensively considers the regulation characteristics of the resources according to the net load prediction result to determine the regulation power of the frequency regulation resources in each future time period, which can improve the deficiencies of real-time distribution. However, this type of method depends on the accuracy of the net load prediction, and too large a prediction error will affect the system frequency quality.

[0007] Before the system runs in real time, the phased allocation method conducts advanced optimal allocation of AGC commands based on the net load prediction results to schedule frequency regulation resources in advance for power adjustment. At the same time, real-time allocation of AGC commands is carried out at the real-time operation moment to meet the actual regulation requirements of the system. This type of method combines the advantages of advanced allocation and real-time allocation, which helps to achieve reasonable allocation of AGC commands.

[0008] In addition, some studies have also applied artificial intelligence technologies such as reinforcement learning to AGC command allocation. To sum up, in order to ensure the frequency quality of high-proportion new energy power systems, domestic and foreign scholars have carried out extensive research on AGC command allocation. However, existing studies have not considered the differences in the response performance of power sources to different AGC commands under different operating conditions. For example, through the analysis of the AGC command response data of a certain gas turbine unit in China, it is found that: under different operating conditions, the response effects of this gas turbine unit to similar AGC commands are quite different (in some conditions, the frequency regulation performance of the unit meets the standards and is not evaluated by the two regulations of the power grid, while in some conditions, the unit is evaluated due to unqualified frequency regulation performance); under similar operating conditions, there are also significant differences in the response effects of this gas turbine unit to different AGC commands. Summary of the Invention

[0009] The object of the present invention is to provide an AGC regulation demand decomposition method that dynamically matches the real-time operating conditions of power sources, including the following steps:

[0010] 1) Construct a power source frequency regulation performance perception model;

[0011] 2) Set the historical frequency regulation performance vector of the power source as the compliance parameter, and input the current multi-dimensional operating conditions of the power source into the power source frequency regulation performance perception model to extract the AGC command adaptation range of the power source;

[0012] 3) Embed the AGC command adaptation range of the power source into the AGC allocation strategy to achieve dynamic matching of AGC command issuance and the operating conditions of the power source.

[0013] Furthermore, in step 1), the power source frequency regulation performance perception model includes a first power source frequency regulation performance perception model and a second power source frequency regulation performance perception model;

[0014] The input of the first power source frequency regulation performance perception model is the multi-dimensional operating conditions of the power source and the up-regulation AGC command, and the output is the power source frequency regulation performance vector at the corresponding moment;

[0015] The input of the second power source frequency regulation performance perception model is the multi-dimensional operating conditions of the power source and the down-regulation AGC command, and the output is the power source frequency regulation performance vector at the corresponding moment;

[0016] The operating conditions include the operating conditions of gas turbine units and the operating conditions of hydroelectric generating units;

[0017] The operating conditions of the gas turbine unit include the current output of the unit, the gas flow control command, the intake air temperature of the air compressor, and the position of the main fuel flow control valve;

[0018] The operating conditions of the hydropower unit include the current output of the unit, the water head, and the guide vane opening.

[0019] Furthermore, the steps for constructing the first power frequency regulation performance perception model include:

[0020] a1) Obtain the historical multi-dimensional operating conditions of the power source, the received historical upward frequency regulation AGC commands, and the historical power frequency regulation performance vector at the corresponding time; the frequency regulation performance vector includes a compliance parameter or a non-compliance parameter; the compliance parameter is 1, and the non-compliance parameter is 0;

[0021] a2) Construct a sample data set with the historical multi-dimensional operating conditions of the power source and the received historical upward frequency regulation AGC commands as inputs and the historical power frequency regulation performance compliance results at the corresponding time as outputs;

[0022] a3) Use the sample data set to train and validate the SVM binary classification model to obtain the first power frequency regulation performance perception model.

[0023] The decision function f(x PCA ) of the SVM binary classification model is as follows:

[0024] f(x PCA ) = w T x PCA + b (1)

[0025] In the formula: w and b respectively represent the normal vector and intercept of the SVM decision hyperplane; x PCA represents the input.

[0026] Furthermore, the steps for constructing the second power frequency regulation performance perception model include:

[0027] b1) Obtain the historical multi-dimensional operating conditions of the power source, the received historical downward frequency regulation AGC commands, and the historical power frequency regulation performance vector at the corresponding time; the frequency regulation performance vector includes a compliance parameter or a non-compliance parameter; the compliance parameter is 1, and the non-compliance parameter is 0;

[0028] b2) Construct a sample data set with the historical multi-dimensional operating conditions of the power source and the received historical downward frequency regulation AGC commands as inputs and the historical power frequency regulation performance compliance results at the corresponding time as outputs;

[0029] b3) Use the sample data set to train and validate the SVM binary classification model to obtain the second power frequency regulation performance perception model.

[0030] Further, when constructing the sample data set, the PCA method is used to extract key features from the historical multi-dimensional operating conditions of the power supply and the received historical AGC commands, and the extracted principal component features are used as the input;

[0031] The principal component features after repeated PCA processing are as follows:

[0032]

[0033] In the formula: k and K are the numbers and sets of the retained principal components respectively; x PCA,k represents the k-th principal component; n and N are the numbers and sets of the dimensions of the original input features respectively; c k,n is the feature contribution rate of the n-th original input feature to the k-th principal component. x n is the feature before processing.

[0034] Further, during the model training process, the update objectives of the normal vector and intercept w and b are as follows:

[0035]

[0036] In the formula: l and L are the numbers and sets of the training samples respectively; ||w|| is the norm of w, which can be expressed as γ l is the slack variable representing the misclassification degree of the sample; C + and C - are the misclassification penalty factors for the samples with qualified and unqualified frequency modulation performance respectively; L + and L - are the sets of samples with qualified and unqualified frequency modulation performance respectively. y l represents the classification result.

[0037] Further, during the model training process, the update process of the normal vector and intercept w and b is as follows:

[0038] c1) Let the penalty factor C +,q = 1, C -,q = 1, and set the maximum number of iterations q max , the initial iteration step size ΔC of the false alarm penalty factor - , and the initial flag f = 0 representing the adjustment method of the false alarm penalty factor;

[0039] c2) Use the sample data set to train the SVM, and calculate the classification precision and classification accuracy in this round of iteration according to the training set, that is:

[0040]

[0041] Wherein, TP and FP are the numbers of true positives and false positives, respectively; TN and FN are the numbers of true negatives and false negatives, respectively.

[0042] c3) If the number of iterations has reached the maximum number of iterations q max , then go to step c8); otherwise, enter step c4).

[0043] c4) If f = 0, adopt the adjustment method of increasing the false alarm penalty factor and go to step c5); otherwise, adopt the adjustment method of decreasing the false alarm penalty factor and go to step c7).

[0044] c5) If indicates that there is no false alarm in the training set, go to step c6); otherwise, let C -,q+1 = C -,q + ΔC - , q = q + 1, and return to step c2).

[0045] c6) If indicates that the increase in the false alarm penalty factor in this round of iteration does not cause a decrease in the classification accuracy, go to step c8); otherwise, it indicates that the increase in the false alarm penalty factor in this round of iteration is too large. Let C -,q+1 = C -,q - ΔC - / 10, q = q + 1, and return to step c2).

[0046] c7) If Let C -,q+1 = C -,q - ΔC - / 10, q = q + 1, and return to step c2); otherwise, it indicates that continuously decreasing the false alarm penalty factor will affect the classification precision. Take the (q - 1)-th round of iteration as the final result and go to step c8).

[0047] c8) Obtain the optimal parameters w * and b * of the SVM model, and output the result.

[0048] Further, the power AGC command adaptation range is as follows:

[0049]

[0050] Wherein: a is the input feature dimension number corresponding to the AGC command; x a is the AGC command received by the power supply. c k,a is the contribution rate.

[0051] Further, the AGC allocation strategy objective function and constraint conditions are as follows, respectively:

[0052]

[0053] where: i is the power supply number; I is the set of power supplies; b A,i is the power supply adjustment cost; A i is the power supply adjustment amount, that is, the AGC command received by the power supply; and are the slack amounts of the upper and lower limits of the power supply command adaptation range; S i,max is the maximum upward / downward adjustable amount of the power supply determined by the adjustment speed and adjustment capacity; M is a constant with a very large value. A i,max 、A i,min are the upper and lower limits of the power supply adjustment amount.

[0054] Furthermore, the steps to achieve the dynamic matching of the AGC command issuance and the power supply operating conditions include:

[0055] d1) Allocate the total system AGC command A sys to each power supply in the order of decreasing adjustment speed; the upper limit of the AGC command received by each power supply is the upper limit A i,max of its adaptation range, and the lower limit is 0;

[0056] d2) Judge whether there are still unallocated commands A sys,re in the system. If there are unallocated commands, it means that only considering the power supply command adaptation range cannot fully meet the system adjustment requirements, and go to step d3); otherwise, go to step d4).

[0057] d3) Considering the actual adjustment upper limit of the power supply determined by the adjustment speed and adjustment capacity, calculate the remaining adjustable amount of each power supply, and allocate the remaining commands A sys,re to each power supply in the order of decreasing remaining adjustable amount, go to step d10), and complete this command allocation.

[0058] d4) Judge whether the command A j received by the last power supply j receiving the command is within its adaptation range [A j,min ,A j,max . If it is within the adaptation range, it means that all power supplies can complete this adjustment task within the command adaptation range, go to step d10), and complete this command allocation; otherwise, go to step d5).

[0059] d5) Judge whether the last power supply j receiving the command is the first power supply with the fastest adjustment speed. If not, go to step d6); otherwise, go to step d7).

[0060] d6) Calculate the difference R j between the command A j,min currently received by the power supply j and the lower limit A j, and calculate the total adjustable amount D that can be reduced by other power supplies receiving instructions under the condition of meeting their own adaptation ranges, and then judge R j and D. If R j is greater than D, go to step d10) to complete this instruction allocation. Otherwise, reduce the instructions received by other power supplies in the order from slowest to fastest adjustment speed. The reduced instructions must not exceed the power supply's own adaptation range, and increase the instructions received by power supply j equally until it reaches the lower limit A j,min of the adaptation range, and go to step d10) to complete this instruction allocation;

[0061] Among them, the total adjustable amount D that can be reduced is as follows:

[0062]

[0063] d7) Judge whether the overall system adjustment requirement is less than the minimum value of the lower limits of all power supply adaptation ranges. If so, do not re-allocate the AGC instructions, go to step d10) to complete this instruction allocation; otherwise, go to step d8) to re-allocate the AGC instructions;

[0064] d8) Screen out the power supplies whose lower limits of the adaptation range are less than the system adjustment requirement, and re-allocate the AGC instructions to these power supplies in the order from fastest to slowest adjustment speed with the upper limit of the power supply adaptation range as the limit;

[0065] d9) Judge whether there are still unallocated instructions in the system. If so, allocate the remaining instructions to other power supplies that have not received instructions in the order from fastest to slowest adjustment speed, and go to step d10) to complete this instruction allocation; otherwise, return to step d4) to adjust the AGC instruction allocation result within the range of the screened power supplies;

[0066] d10) Obtain the instruction allocation results of all power supplies and send them to each power supply.

[0067] The technical effect of the present invention is beyond doubt. The present invention proposes an AGC regulation demand decomposition strategy for dynamically matching the real-time operating conditions of power supplies. By using machine learning technology, the instruction range that the power supply can effectively respond to under the current operating conditions is extracted and embedded in the AGC allocation, thereby improving the matching degree between the AGC instructions and the power supply operating state. The proposed method can be used not only for the allocation of the overall grid AGC instructions, but also for scenarios such as the decomposition of AGC instructions from a virtual power plant to internal regulation resources.

[0068] The beneficial effects of the present invention are as follows:

[0069] 1) A method for extracting the adaptation range of power commands based on frequency modulation performance perception is proposed. The operating conditions of the power supply and the received AGC commands are used as input features, and principal component analysis (PCA) is adopted to extract key input information; a perception model for meeting the frequency modulation performance standards of the power supply is constructed based on the support vector machine (SVM), and an adaptive penalty factor is introduced to address false alarm situations to improve the perception accuracy; under given operating conditions, the command range that the power supply can effectively respond to is deduced according to the SVM perception model.

[0070] 2) An AGC allocation strategy incorporating the adaptation range of power commands is proposed. For lead optimization allocation, a soft constraint oriented to the adaptation range of power commands is constructed; for real-time allocation, an allocation rule considering the adaptation range of power commands is designed. By incorporating the adaptation range of power commands, the real-time operating conditions of the power supply can be fully considered during the AGC command allocation process, thereby improving the adaptability of AGC commands to the operating state of the power supply under the condition of meeting the overall regulation requirements of the system.

[0071] From the perspective of the actual operation of the power grid, the extraction of the adaptation range of power commands can be carried out on the power supply side. The power supply side only transmits the command adaptation range to the grid side as the basis for its execution of AGC command allocation, and the power supply side does not need to transmit a large amount of internal data to the grid side.

[0072] The AGC allocation strategy incorporating the adaptation range of power commands allocates adjustment commands that match the operating conditions of each power supply on the premise of meeting the system regulation requirements, thereby ensuring the command response performance of the power supply and improving the system frequency quality. Description of the Drawings

[0073] Figure 1 is the extraction framework for the adaptation range of power commands based on frequency modulation performance perception;

[0074] Figure 2 is the iterative search process for the optimal false alarm penalty factor;

[0075] Figure 3 is the AGC real-time allocation process incorporating the adaptation range of power commands;

[0076] Figure 4 is the comparison of the adjustment process of the false alarm penalty factors of M3 and M4 in the down-frequency modulation performance perception model. Detailed Implementation Manner

[0077] The present invention will be further described below in conjunction with embodiments, but it should not be understood that the above-mentioned subject scope of the present invention is limited to the following embodiments. Without departing from the above technical idea of the present invention, various substitutions and changes made according to the common general knowledge and customary means in the art shall be included within the protection scope of the present invention.

[0078] Embodiment 1:

[0079] Refer to Figures 1 to 4 , an AGC regulation demand decomposition method for dynamically matching the real-time operating conditions of power sources, comprising the following steps:

[0080] 1) Construct a power frequency regulation performance perception model;

[0081] 2) Set the power source historical frequency regulation performance vector as the compliance parameter, and input the current multi-dimensional operating conditions of the power source into the power frequency regulation performance perception model to extract the AGC command adaptation range of the power source;

[0082] 3) Embed the AGC command adaptation range of the power source into the AGC allocation strategy to achieve dynamic matching of AGC command issuance and power source operating conditions.

[0083] In step 1), the power frequency regulation performance perception model includes a first power frequency regulation performance perception model and a second power frequency regulation performance perception model;

[0084] The input of the first power frequency regulation performance perception model is the multi-dimensional operating conditions of the power source and the up-regulation AGC command, and the output is the power frequency regulation performance vector at the corresponding moment;

[0085] The input of the second power frequency regulation performance perception model is the multi-dimensional operating conditions of the power source and the down-regulation AGC command, and the output is the power frequency regulation performance vector at the corresponding moment;

[0086] The operating conditions include the operating conditions of gas turbine units and hydroelectric generating units;

[0087] The operating conditions of the gas turbine unit include the current output of the unit, the gas flow control command, the intake air temperature of the air compressor, and the position of the main fuel flow control valve;

[0088] The operating conditions of the hydroelectric generating unit include the current output of the unit, the water head, and the guide vane opening;

[0089] The steps for constructing the first power frequency regulation performance perception model include:

[0090] a1) Obtain the historical multi-dimensional operating conditions of the power source, the received historical up-regulation AGC command, and the power source historical frequency regulation performance vector at the corresponding moment; the frequency regulation performance vector includes compliance parameters or non-compliance parameters; the compliance parameter is 1, and the non-compliance parameter is 0;

[0091] a2) Construct a sample data set with the multi-dimensional operating conditions of the power supply history and the received historical upward frequency modulation AGC commands as inputs, and the power supply historical frequency modulation performance compliance results at corresponding times as outputs;

[0092] a3) Use the sample data set to train and validate the SVM binary classification model to obtain the first power supply frequency modulation performance perception model.

[0093] The decision function f(x PCA ) of the SVM binary classification model is as follows:

[0094] f(x PCA ) = w T x PCA + b (1)

[0095] In the formula: w and b respectively represent the normal vector and intercept of the SVM decision hyperplane; x PCA represents the input.

[0096] The steps to construct the second power supply frequency modulation performance perception model include:

[0097] b1) Obtain the multi-dimensional operating conditions of the power supply history, the received historical downward frequency modulation AGC commands, and the power supply historical frequency modulation performance vector at corresponding times; the frequency modulation performance vector includes compliance parameters or non-compliance parameters; the compliance parameter is 1, and the non-compliance parameter is 0;

[0098] b2) Construct a sample data set with the multi-dimensional operating conditions of the power supply history and the received historical downward frequency modulation AGC commands as inputs, and the power supply historical frequency modulation performance compliance results at corresponding times as outputs;

[0099] b3) Use the sample data set to train and validate the SVM binary classification model to obtain the second power supply frequency modulation performance perception model.

[0100] When constructing the sample data set, use the PCA method to extract key features from the multi-dimensional operating conditions of the power supply history and the received historical AGC commands, and use the extracted principal component features as inputs;

[0101] The principal component features after repeated PCA processing are as follows:

[0102]

[0103] In the formula: k and K are respectively the retained principal component numbers and sets; x PCA,k represents the k-th principal component; n and N are respectively the dimension numbers and sets of the original input features; c k,n is the feature contribution rate of the n-th original input feature to the k-th principal component. x n is the feature before processing.

[0104] During the model training process, the update objectives of the normal vector and the intercepts w and b are as follows:

[0105]

[0106] where: l and L are the numbers and sets of training samples respectively; ||w|| is the norm of w, which can be expressed as γ l is a slack variable representing the misclassification degree of the sample; C + and C - are the misclassification penalty factors for samples with qualified and unqualified FM performance respectively; L + and L - are the sets of samples with qualified and unqualified FM performance respectively. y l represents the classification result.

[0107] During the model training process, the update process of the normal vector and the intercepts w and b is as follows:

[0108] c1) Let the penalty factor C +,q = 1, C -,q = 1, and set the maximum number of iterations q max , the initial iteration step size ΔC - of the false alarm penalty factor, and the initial flag f = 0 representing the adjustment method of the false alarm penalty factor;

[0109] c2) Train the SVM using the sample dataset, and calculate the classification precision and classification accuracy in this round of iteration based on the training set, that is:

[0110]

[0111] where TP and FP are the numbers of true positives and false positives; TN and FN are the numbers of true negatives and false negatives;

[0112] c3) If the number of iterations has reached the maximum number of iterations q max , go to step c8), otherwise, enter step c4);

[0113] c4) If f = 0, adopt the adjustment method of increasing the false alarm penalty factor and go to step c5); otherwise, adopt the adjustment method of decreasing the false alarm penalty factor and go to step c7);

[0114] c5) If indicates that there is no false alarm in the training set, go to step c6), otherwise, let C -,q+1 = C -,q + ΔC - , q = q + 1, and return to step c2);

[0115] c6) If It shows that the increase in the false alarm penalty factor in this iteration does not cause a decrease in the classification accuracy. Go to step c8). Otherwise, it shows that the increase in the false alarm penalty factor in this iteration is too large. Let C -,q+1 = C -,q - ΔC - / 10, q = q + 1, and return to step c2);

[0116] c7) If Let C -,q+1 = C -,q - ΔC - / 10, q = q + 1, and return to step c2); Otherwise, it shows that continuously reducing the false alarm penalty factor will affect the classification precision. Take the (q - 1)-th iteration as the final result and go to step c8);

[0117] c8) Obtain the optimal parameters w * and b * of the SVM model, and output the result.

[0118] The power AGC command adaptation range is as follows:

[0119]

[0120] In the formula: a is the input feature dimension number corresponding to the AGC command; x a is the AGC command received by the power supply. c k,a is the contribution rate.

[0121] The AGC allocation strategy objective function and constraint conditions are as follows respectively:

[0122]

[0123] In the formula: i is the power supply number; I is the power supply set; b A,i is the power supply adjustment cost; A i is the power supply adjustment amount, that is, the AGC command received by the power supply; and are the relaxation amounts of the upper and lower limits of the power supply command adaptation range; S i,max is the maximum upward / downward adjustable amount of the power supply determined by the adjustment speed and adjustment capacity; M is a constant with a very large value. A i,max and A i,min are the upper and lower limits of the power supply adjustment amount.

[0124] The steps to achieve the dynamic matching of the AGC command issuance and the power supply operation conditions include:

[0125] d1) Arrange the total system AGC command A sysAllocated to each power supply; the upper limit of the AGC command received by each power supply is the upper limit A of its adaptation range i,max , and the lower limit is 0;

[0126] d2) Determine whether there are still unallocated commands in the system A sys,re , if there are unallocated commands, it indicates that only considering the power command adaptation range cannot fully meet the system regulation requirements, go to step d3); otherwise, go to step d4).

[0127] d3) Considering the actual regulation upper limit of the power supply determined by the regulation speed and regulation capacity, calculate the remaining adjustable amount of each power supply, and allocate the remaining commands A sys,re to each power supply in descending order of the remaining adjustable amount, go to step d10), and complete this command allocation.

[0128] d4) Determine whether the command received by the power supply j that receives the last command A j is within its adaptation range [A j,min , A j,max , if it is within the adaptation range, it indicates that all power supplies can complete this regulation task within the command adaptation range, go to step d10), and complete this command allocation; otherwise, go to step d5).

[0129] d5) Determine whether the power supply j that receives the last command is the first power supply with the fastest regulation speed. If not, go to step d6); otherwise, go to step d7).

[0130] d6) Calculate the difference R j between the command currently received by the power supply j A j,min and the lower limit of the adaptation range A j , and calculate the total adjustable reduction D that other power supplies that have received commands can reduce under the condition of meeting their own adaptation range. Then judge the size of R j and D. If R j is greater than D, go to step d10) to complete this command allocation. Otherwise, reduce the commands received by other power supplies in ascending order of regulation speed. The reduced commands must not exceed the power supply's own adaptation range, and increase the commands received by the power supply j equally until it reaches the lower limit of the adaptation range A j,min , go to step d10) to complete this command allocation;

[0131] Among them, the total adjustable reduction is as follows:

[0132]

[0133] d7) Determine whether the overall system regulation demand is less than the minimum value of the lower limits of all power supply adaptation ranges. If so, do not reallocate the AGC command, go to step d10), and complete this command allocation; otherwise, go to step d8) to reallocate the AGC command;

[0134] d8) Screen out the power supplies whose lower limits of the adaptation range are less than the system regulation demand, and reallocate the AGC command to these power supplies in descending order of the regulation speed with the upper limit of the power supply adaptation range as the limit;

[0135] d9) Determine whether there are still unallocated commands in the system. If so, allocate the remaining commands to other power supplies that have not received commands in descending order of the regulation speed, and go to step d10) to complete this command allocation; otherwise, return to step d4) to adjust the AGC command allocation result within the range of the screened power supplies;

[0136] d10) Obtain the command allocation results of all power supplies and send them to each power supply.

[0137] Embodiment 2:

[0138] The AGC regulation demand decomposition method for dynamically matching the real-time operating conditions of power supplies includes the following steps:

[0139] 1) Construct a power frequency modulation performance perception model;

[0140] 2) Set the power historical frequency modulation performance vector as the compliance parameter, and input the current multi-dimensional operating conditions of the power supply into the power frequency modulation performance perception model to extract the AGC command adaptation range of the power supply;

[0141] 3) Embed the AGC command adaptation range of the power supply into the AGC allocation strategy to achieve dynamic matching of the AGC command issuance and the operating conditions of the power supply.

[0142] Embodiment 3:

[0143] The AGC regulation demand decomposition method for dynamically matching the real-time operating conditions of power supplies has the same technical content as Embodiment 2. Further, in step 1), the power frequency modulation performance perception model includes a first power frequency modulation performance perception model and a second power frequency modulation performance perception model;

[0144] The input of the first power frequency modulation performance perception model is the multi-dimensional operating conditions of the power supply and the up-regulation AGC command, and the output is the power frequency modulation performance vector at the corresponding moment;

[0145] The input of the first power frequency modulation performance perception model is the multi-dimensional operating conditions of the power supply and the down-regulation AGC command, and the output is the power frequency modulation performance vector at the corresponding moment;

[0146] The operating conditions include the operating conditions of gas turbine units and the operating conditions of hydroelectric generating units;

[0147] The operating conditions of the gas turbine unit include the current output of the unit, the gas flow control command, the intake air temperature of the air compressor, and the position of the main fuel flow control valve;

[0148] The operating conditions of the hydropower unit include the current output of the unit, the water head, and the guide vane opening;

[0149] Embodiment 4:

[0150] The AGC regulation demand decomposition method for dynamically matching the real-time operating conditions of the power source, the technical content is the same as any one of Embodiments 2-3. Further, the steps of constructing the first power source frequency modulation performance perception model include:

[0151] a1) Obtain the historical multi-dimensional operating conditions of the power source, the historical up-regulation frequency modulation AGC commands received, and the historical power source frequency modulation performance vectors at corresponding times; the frequency modulation performance vectors include compliance parameters or non-compliance parameters; the compliance parameter is 1, and the non-compliance parameter is 0;

[0152] a2) Construct a sample data set with the historical multi-dimensional operating conditions of the power source and the historical up-regulation frequency modulation AGC commands received as inputs and the historical power source frequency modulation performance compliance results at corresponding times as outputs;

[0153] a3) Use the sample data set to train and verify the SVM binary classification model to obtain the first power source frequency modulation performance perception model.

[0154] The decision function of the SVM binary classification model is as follows:

[0155] f(x PCA ) = w T x PCA + b (1)

[0156] In the formula: w and b respectively represent the normal vector and intercept of the SVM decision hyperplane.

[0157] Embodiment 5:

[0158] The AGC regulation demand decomposition method for dynamically matching the real-time operating conditions of the power source, the technical content is the same as any one of Embodiments 2-4. Further, the steps of constructing the second power source frequency modulation performance perception model include:

[0159] b1) Obtain the historical multi-dimensional operating conditions of the power source, the historical down-regulation frequency modulation AGC commands received, and the historical power source frequency modulation performance vectors at corresponding times; the frequency modulation performance vectors include compliance parameters or non-compliance parameters; the compliance parameter is 1, and the non-compliance parameter is 0;

[0160] b2) Construct a sample data set with the multi-dimensional operating conditions of the power supply history and the received historical down-regulation AGC commands as inputs, and the power supply historical frequency regulation performance compliance results at the corresponding moments as outputs;

[0161] b3) Use the sample data set to train and validate the SVM binary classification model to obtain the second power supply frequency regulation performance perception model.

[0162] Example 6:

[0163] The AGC adjustment demand decomposition method for dynamically matching the real-time operating conditions of the power supply has the same technical content as any one of Examples 2-5. Further, when constructing the sample data set, the PCA method is used to extract key features from the multi-dimensional operating conditions of the power supply history and the received historical AGC commands, and the extracted principal component features are used as inputs;

[0164] The principal component features after repeated PCA processing are as follows:

[0165]

[0166] In the formula: k and K are the numbers and sets of the retained principal components respectively; x PCA,k represents the k-th principal component; n and N are the numbers and sets of the dimensions of the original input features respectively; c k,n is the feature contribution rate of the n-th original input feature to the k-th principal component.

[0167] Example 7:

[0168] The AGC adjustment demand decomposition method for dynamically matching the real-time operating conditions of the power supply has the same technical content as any one of Examples 2-6. Further, during the model training process, the update objectives of the normal vector and the intercept w and b are as follows:

[0169]

[0170] In the formula: l and L are the numbers and sets of the training samples respectively; ||w|| is the norm of w, which can be expressed as γ l is the slack variable representing the misclassification degree of the sample; C + and C - are the misclassification penalty factors for the frequency regulation performance compliance samples and non-compliance samples respectively; L + and L - are the sets of the frequency regulation performance compliance samples and non-compliance samples respectively. y l represents the classification result.

[0171] Example 8:

[0172] AGC regulation demand decomposition method for dynamically matching the real-time operating conditions of the power supply, the technical content is the same as any one of Embodiments 2-7. Further, during the model training process, the update process of the normal vector and the intercepts w and b is as follows:

[0173] c1) Let the penalty factor C +,q = 1, C -,q = 1, and set the maximum number of iterations q max , the initial iteration step size ΔC of the false alarm penalty factor - , and the initial flag f representing the adjustment method of the false alarm penalty factor = 0;

[0174] c2) Train the SVM using the sample data set, and calculate the classification precision and classification accuracy in this round of iteration according to the training set, that is:

[0175]

[0176] c3) If the number of iterations has reached the maximum number of iterations q max , then go to step c8), otherwise, enter step c4);

[0177] c4) If f = 0, adopt the adjustment method of increasing the false alarm penalty factor and go to step c5); otherwise, adopt the adjustment method of decreasing the false alarm penalty factor and go to step c7);

[0178] c5) If indicates that there is no false alarm situation in the training set, go to step c6), otherwise, let C -,q+1 = C -,q + ΔC - , q = q + 1, and return to step c2);

[0179] c6) If indicates that the increase in the false alarm penalty factor in this round of iteration does not cause a decrease in the classification accuracy, go to step c8), otherwise, it indicates that the increase in the false alarm penalty factor in this round of iteration is too large, let C -,q+1 = C -,q - ΔC - / 10, q = q + 1, and return to step c2);

[0180] c7) If let C -,q+1 = C -,q - ΔC - / 10, q = q + 1, and return to step c2); otherwise, it indicates that continuing to decrease the false alarm penalty factor will affect the classification precision, take the (q - 1)-th round of iteration as the final result and go to step c8);

[0181] c8) Obtain the optimal parameters w of the SVM model* and b * , the output result.

[0182] Example 9:

[0183] The AGC regulation demand decomposition method for dynamically matching the real-time operating conditions of the power supply, the technical content is the same as any one of Examples 2-8. Further, the AGC instruction adaptation range of the power supply is as follows:

[0184]

[0185] In the formula: a is the input feature dimension number corresponding to the AGC instruction; x a is the AGC instruction received by the power supply.

[0186] Example 10:

[0187] The AGC regulation demand decomposition method for dynamically matching the real-time operating conditions of the power supply, the technical content is the same as any one of Examples 2-9. Further, the objective function and constraint conditions of the AGC allocation strategy are as follows:

[0188]

[0189] In the formula: i is the power supply number; I is the power supply set; b A,i is the power supply regulation cost; A i is the power supply regulation amount, that is, the AGC instruction received by the power supply; and are the slack amounts of the upper and lower limits of the power supply instruction adaptation range; S i,max is the maximum upward / downward adjustable amount of the power supply determined by the regulation speed and regulation capacity; M is a constant with a very large value.

[0190] Example 11:

[0191] The AGC regulation demand decomposition method for dynamically matching the real-time operating conditions of the power supply, the technical content is the same as any one of Examples 2-10. Further, the steps to realize the dynamic matching of the AGC instruction issuance and the power supply operating conditions include:

[0192] d1) Allocate the total system AGC instruction A sys to each power supply in the order of decreasing regulation speed; the upper limit of the AGC instruction received by each power supply is the upper limit A i,max of its adaptation range, and the lower limit is 0;

[0193] d2) Judge whether there is still an unallocated instruction A sys,re in the system. If there is an unallocated instruction, it means that only considering the power supply instruction adaptation range cannot fully meet the system regulation demand, and go to step d3); otherwise, go to step d4).

[0194] d3) Consider the actual regulation upper limit of the power supply determined by the regulation speed and regulation capacity, calculate the remaining adjustable amount of each power supply, and allocate the remaining command A in descending order of the remaining adjustable amount to each power supply, then go to step d10) to complete this command allocation. sys,re Allocate to each power supply, go to step d10), and complete this command allocation.

[0195] d4) Judge whether the command A received by the power supply j that receives the last command is within its adaptation range [A j , A j,min , A j,max . If it is within the adaptation range, it means that all power supplies can complete this regulation task within the command adaptation range. Go to step d10) to complete this command allocation; otherwise, go to step d5).

[0196] d5) Judge whether the power supply j that receives the last command is the first power supply with the fastest regulation speed. If not, go to step d6); otherwise, go to step d7).

[0197] d6) Calculate the difference R between the command A currently received by the power supply j and the lower limit A of the adaptation range j and calculate the total adjustable amount D that can be reduced by other power supplies that have received commands under the condition of meeting their own adaptation ranges. Then judge the size of R j,min and D. If R j is greater than D, go to step d10) to complete this command allocation. Otherwise, reduce the commands received by other power supplies in ascending order of the regulation speed. The reduced commands must not exceed the own adaptation range of the power supply, and increase the command received by the power supply j equally until it reaches the lower limit A j of the adaptation range, then go to step d10) to complete this command allocation; j where the total adjustable amount D that can be reduced is as follows: j,min Go to step d10) to complete this command allocation;

[0198] Among them, the total adjustable amount D that can be reduced is as follows:

[0199]

[0200] d7) Judge whether the overall system regulation demand is less than the minimum value of the lower limits of the adaptation ranges of all power supplies. If so, do not re-allocate the AGC command, go to step d10) to complete this command allocation; otherwise, go to step d8) to re-allocate the AGC command;

[0201] d8) Screen out the power supplies whose lower limits of the adaptation ranges are less than the system regulation demand, and re-allocate the AGC commands to these power supplies with the upper limits of the adaptation ranges as the limits in descending order of the regulation speed;

[0202] d9) Determine whether there are unallocated instructions in the system. If so, allocate the remaining instructions to the other power supplies that have not received instructions in the order of decreasing adjustment speed, and go to step d10) to complete this instruction allocation; otherwise, return to step d4) and adjust the AGC instruction allocation result within the selected power supply range.

[0203] d10) Obtain the instruction allocation results of all power supplies and send them to each power supply.

[0204] Embodiment 12:

[0205] The AGC adjustment demand decomposition method for dynamically matching the real-time operating conditions of power supplies includes the following steps:

[0206] 1. Extraction of the power supply instruction adaptation range based on frequency modulation performance perception

[0207] The framework for extracting the power supply instruction adaptation range based on frequency modulation performance perception proposed by the present invention is as Figure 1 shown.

[0208] The proposed framework includes two stages, which are described as follows:

[0209] Stage 1 is the construction of the power supply frequency modulation performance perception model. In the present invention, the historical multi-dimensional operating conditions of the power supply and the received historical AGC instructions are used as sample inputs, and whether the power supply historical frequency modulation performance meets the standard at the corresponding moment is used as the sample output. A frequency modulation performance compliance perception model based on SVM is constructed, which is a binary classification model. Among them, PCA is used to extract key information from multi-dimensional input features to improve the accuracy of frequency modulation performance perception; an adaptive penalty factor for false alarm situations is embedded to improve the precision of frequency modulation performance perception, avoid classifying samples with non-compliant frequency modulation performance as compliant, and ensure the credibility of the subsequent instruction adaptation range extraction results. After sample training, the power supply frequency modulation performance compliance rules can be obtained through the constructed SVM model, laying a foundation for extracting the AGC instruction range that the power supply can effectively respond to under the given operating conditions. In addition, during the actual operation of the system, the latest power supply operation data can be continuously added to the training data to ensure the accuracy of power supply frequency modulation performance perception through rolling update of training samples.

[0210] Stage 2 is the extraction of the power supply instruction adaptation range. During the actual operation process, based on the power supply frequency modulation performance compliance rules obtained by the SVM model, the instruction range that can make its frequency modulation performance meet the standard is deduced according to the current operating conditions of the power supply. Subsequently, combined with operating parameters such as the power supply adjustment speed and adjustment capacity, the final power supply instruction adaptation range is obtained.

[0211] The power supply frequency modulation performance perception model (Stage 1) and the instruction adaptation range extraction method (Stage 2) proposed by the present invention will be introduced in detail below.

[0212] The present invention uses the multi-dimensional operating condition data of the power supply and the AGC command data as the sample input feature x. Among them, the collected operating conditions of the gas turbine units include 10 types of data such as the current output of the unit, the gas flow control command, the intake air temperature of the air compressor, and the position of the main fuel flow control valve. The operating conditions of the hydropower units include 3 types of data such as the current output of the unit, the water head, and the guide vane opening; the AGC command data includes the frequency increase command and the frequency decrease command. The response effects of the power supply to different-direction AGC commands under the same operating conditions may vary. Therefore, the present invention establishes different frequency modulation performance perception models for the frequency increase command and the frequency decrease command respectively, and the AGC command data in the input feature only reflects the command size. Whether the frequency modulation performance of the power supply meets the standard is used as the output feature y, where -1 indicates that the frequency modulation performance of the power supply does not meet the standard, and +1 indicates that the frequency modulation performance of the power supply meets the standard.

[0213] There may be redundant information in the above multi-dimensional input features, which will affect the accuracy of frequency modulation performance perception. Therefore, the present invention uses PCA to extract the key information in the original input features. After PCA processing, the principal components to be retained are selected in order from largest to smallest contribution until the cumulative contribution exceeds a certain threshold (set to 95% in the present invention). The extracted principal components are composed of the original variables and reflect most of the information provided by the original variables. The principal components retained after PCA processing can be expressed as:

[0214]

[0215] In the formula: k and K are the serial numbers and sets of the retained principal components respectively; x PCA,k represents the k-th principal component; n and N are the serial numbers and sets of the dimensions of the original input features respectively; c k,n is the characteristic contribution rate of the n-th original input feature to the k-th principal component.

[0216] The retained principal components are used as the final sample input, thus forming a binary classification sample set (x PCA , y) for constructing the frequency modulation performance perception model.

[0217] The problem of frequency modulation performance perception is a typical binary classification problem. SVM has a high fitting accuracy for binary classification problems under small sample conditions, and can achieve an explicit analytical expression of the frequency modulation performance compliance rule when using a linear kernel function. Therefore, the present invention uses SVM to construct the power supply frequency modulation performance perception model.

[0218] The SVM decision function using a linear kernel function can be expressed as:

[0219] f(x PCA ) = w T x PCA + b (2)

[0220] Where: w and b are the parameters of SVM, representing the normal vector and intercept of the SVM decision hyperplane respectively.

[0221] The classification method of SVM is: If formula (2) is greater than or equal to 0, the sample is considered to be a positive sample (that is, the frequency modulation performance meets the standard); otherwise, the sample is a negative sample.

[0222] The process of obtaining the optimal parameters w and b of SVM based on training samples can be expressed as the following optimization problem:

[0223]

[0224] Where: l and L are the number and set of training samples respectively; ||w|| is the norm of w, which can be expressed as γ l C is a slack variable that indicates the degree of sample misclassification. Since samples are generally not strictly linearly separable, some samples are usually allowed to be misclassified during the practical application of SVM. l is the misclassification penalty factor.

[0225] The classification error of the SVM model is inevitable. For the binary classification problem involved in the present invention, the misclassified samples mainly include two situations: samples whose actual frequency modulation performance meets the standard but are classified as substandard (i.e., false negative examples, also known as missed reports), and samples whose actual frequency modulation performance does not meet the standard but are classified as meeting the standard (i.e., false positive examples, also known as false reports). Among them, the missed reports will cause the subsequent power supply instruction adaptation range to be unable to cover all adaptation instructions, but based on this range, there will be no situation where incompatible instructions are sent to the power supply; and the false report situation will cause incompatible instructions to be included in the adaptation range. If based on this range, there may be a situation where incompatible instructions are sent to the power supply, affecting the power supply frequency modulation performance and system frequency quality. Therefore, the frequency modulation performance perception model constructed by the present invention should try to avoid the occurrence of false reports, which is manifested in that the accuracy of the classification result (as shown in formula (4)) should be as high as possible.

[0226]

[0227] Where: P P is the classification accuracy; TP is the number of samples whose actual FM performance meets the standard and are also classified as meeting the standard; FP is the number of samples whose actual FM performance does not meet the standard but are classified as meeting the standard.

[0228] To this end, the present invention introduces an adaptive false alarm penalty factor to increase the penalty for false alarms according to the classification results of the training set to improve the classification accuracy, which is specifically introduced as follows.

[0229] The objective function of the optimization problem shown in formula (3) can be rewritten as:

[0230]

[0231] Where: C + and C - are the misclassification penalty factors for the samples with qualified and unqualified FM performance respectively; L + and L - are the sets of samples with qualified and unqualified FM performance respectively. To increase the penalty for false alarms, it is necessary to satisfy C + ≤C - .

[0232] Increasing C - helps to improve the classification precision and reduce false alarms, but an excessive C - will lead to an increase in missed alarms, affecting the classification accuracy (as shown in Equation (6)).

[0233]

[0234] Where: P A is the classification accuracy; TN is the number of samples with unqualified FM performance that are correctly classified; FN is the number of samples with unqualified FM performance that are incorrectly classified.

[0235] Therefore, the present invention uses the classification precision and classification accuracy of the training set to guide the iterative optimization of C - to obtain a false alarm penalty factor that can make the classification precision of the training set reach 100% without overly affecting the classification accuracy. The process is as Figure 2 shown. The main idea is: first, use a relatively large increase step size to quickly obtain a false alarm penalty factor that can make the classification precision reach 100%, and then use a relatively small decrease step size to fine-tune the false alarm penalty factor to alleviate its impact on the classification accuracy, so as to quickly find the most suitable false alarm penalty factor.

[0236] The specific steps are as follows:

[0237] 1) Set the initial values of the penalty factors: C +,q = 1 and C -,q = 1 (q represents the number of iterations), and set the maximum number of iterations q max , the initial iteration step size ΔC of the false alarm penalty factor - (ΔC - > 0) and the initial flag representing the adjustment method of the false alarm penalty factor (denoted as f = 0);

[0238] 2) Train the SVM using the training set and calculate the classification precision P P q and classification accuracy of this round of iteration according to the training set

[0239] 3) If the number of iterations has reached the maximum number of iterations q max , go to step 8);

[0240] 4) If f = 0, adopt the adjustment method of increasing the false alarm penalty factor and go to step 5); otherwise, adopt the adjustment method of decreasing the false alarm penalty factor and go to step 7);

[0241] 5) If indicates that there is no false alarm situation in the training set, go to step 6); otherwise, let C -,q+1 = C -,q + ΔC - , q = q + 1, and return to step 2);

[0242] 6) If indicates that the increase in the false alarm penalty factor in this round of iteration does not cause a decrease in the classification accuracy, go to step 8); otherwise, it indicates that the increase in the false alarm penalty factor in this round of iteration may be too large. The adjustment method of decreasing the false alarm penalty factor (adjust the flag f to 1) will be adopted to find the false alarm penalty factor that can exactly make the classification precision reach 100%, so as to alleviate the impact of the increase in the false alarm penalty factor on the classification accuracy . Let C -,q+1 = C -,q - ΔC - / 10, q = q + 1, and return to step 2);

[0243] 7) If Let C -,q+1 = C -,q - ΔC - / 10, q = q + 1, and return to step 2); otherwise, it indicates that continuing to decrease the false alarm penalty factor will affect the classification precision. Take the (q - 1)-th round of iteration as the final result and go to step 8);

[0244] 8) Obtain the optimal parameters w * and b * of the SVM model and output the result.

[0245] Compared with directly adopting the adjustment method of a smaller increase step of the false alarm penalty factor, the proposed adjustment method can more efficiently find the most suitable false alarm penalty factor, reserving enough time for the subsequent extraction of the instruction adaptation range and the allocation of AGC instructions.

[0246] After training, under the given operating conditions and AGC instructions, the condition for the power frequency modulation performance to meet the standard can be expressed as:

[0247] w *T x PCA + b * ≥ 0 (7)

[0248] where: w * and b * are the optimal parameters of the SVM after training.

[0249] Based on this, taking the power operation condition as the known quantity and the AGC command as the unknown quantity, it can be inversely deduced that in order to meet the power frequency regulation performance standard, the conditions that the AGC command received by the power supply should satisfy. Therefore, first substitute Equation (1) into Equation (7) and expand it to obtain the frequency regulation performance standard conditions related to the original input features:

[0250]

[0251] By transforming Equation (8), the power command adaptation range shown in Equation (9) can be obtained.

[0252]

[0253] where: a is the input feature dimension number corresponding to the AGC command; x a is the AGC command received by the power supply. In Equation (9), the direction of the inequality sign needs to be adjusted according to the positive or negative value of the denominator of the right - hand side term. Here, the expression when the denominator is positive is given.

[0254] The adjustable power that the power supply can respond to is limited by its own adjustment speed and adjustment capacity. Combining with the power command adaptation condition given in Equation (9), the command adaptation range that the power supply can respond to can be obtained (denoted as [A min , A max ). If the response range of the power supply conflicts with the command adaptation condition, it indicates that the adjustment performance of the power supply is poor under the current working condition. At this time, there is no adapted command for the power supply (which can be denoted as [0, 0]).

[0255] As mentioned above, affected by the false alarm situation, the present invention cannot guarantee that the power supply can effectively respond to the AGC commands belonging to the extracted adaptation range. However, compared with the traditional AGC allocation strategy that does not consider the adaptation range (believing that the power supply can effectively respond to the commands that meet its own adjustment limit), even if an incompletely accurate adaptation range is embedded, it will not damage the command allocation effect. In addition, the introduction of the adaptive false alarm penalty factor can effectively reduce the false alarm rate and ensure that the power supply can effectively respond to the AGC commands within the extracted adaptation range in most scenarios.

[0256] In summary, as the power operation condition changes continuously, only by inputting its real - time operation condition, the method proposed by the present invention can extract the command adaptation range of the power supply under the corresponding working condition, thereby providing a basis for the allocation of grid AGC commands and improving the matching degree between AGC command allocation and the real - time operation condition of the power supply.

[0257] In addition, it should be noted that the current modes for the power source to respond to AGC commands mainly include the following two: 1) Multiple generating units within the power source each receive the grid AGC command and are separately evaluated by the grid; 2) The power source receives the grid AGC command as a whole, and then further decomposes the total command to each internal generating unit. The power source will be evaluated by the grid as a whole. For Mode 1, the power source can separately extract the command adaptation range of each unit. As mentioned above, the actual gas turbine unit mentioned in the present invention belongs to this mode; for Mode 2, the power source can integrate the operating condition data of each internal unit as the input feature for extracting the command adaptation range, so as to extract the command adaptation range for the power source as a whole. The actual hydroelectric generating unit mentioned in the present invention belongs to this mode. As long as the way of decomposing the total command to each internal unit of the power source does not change, the power source frequency modulation performance perception model constructed based on historical data can reflect the correlation between the operating conditions of each internal unit, the total AGC command of the power source, and the overall frequency modulation performance of the power source. However, when the total command decomposition method changes, it is necessary to collect the power source operation data again and construct a power source frequency modulation performance perception model. The results of the numerical example simulation show that the proposed methods in both modes can accurately perceive the power source frequency modulation performance, verifying the applicability of the methods in both modes.

[0258] 2 AGC Allocation Strategy with Embedded Power Source Command Adaptation Range

[0259] In this section, the extracted power source command adaptation range will be embedded in the AGC allocation strategy to achieve the dynamic matching of AGC command issuance and the operating conditions of the power source. As mentioned in the introduction, the currently commonly used AGC command allocation strategies include AGC advanced optimization allocation and AGC real-time allocation. Taking the up-frequency modulation AGC command as an example, the embedding methods of the power source command adaptation range in the two allocation strategies will be elaborated below. The present invention mainly focuses on how to embed the power source command adaptation range in the commonly used allocation strategies, rather than comparing the advantages and disadvantages of the advanced allocation strategy and the real-time allocation strategy.

[0260] The AGC advanced optimization allocation strategy constructs an optimization allocation model based on the net load prediction result to perform the optimal allocation of AGC commands, and its objective function is generally to minimize the regulation cost, as shown in Equation (10).

[0261]

[0262] where: i is the power source number; I is the power source set; b A,i is the power source regulation cost; A i is the power source regulation amount, that is, the AGC command received by the power source.

[0263] The operation constraints include regulation demand constraints, regulation capacity constraints, power source ramp constraints, etc.

[0264] To ensure that the AGC commands received by the power sources are within their adaptation ranges, the most direct approach is to embed the extracted adaptation range of the power commands [A min , A max as a constraint into the optimization allocation model. However, compared with the adjustable range of the power sources determined by the adjustment speed and adjustment capacity, the adaptation range of the power commands may be smaller, and it may not be able to effectively meet the system adjustment requirements under the condition of considering all the adaptation ranges of the power commands. For this reason, taking the allocation of the frequency regulation AGC commands as an example, the present invention constructs the adaptation range of the power commands as soft constraints shown in (11)-(13), and adds a slack penalty term (shown in Equation (14)) to the objective function, so as to consider the adaptation range of the power commands only when the system adjustment requirements are met (the adjustment requirement constraint is a hard constraint that must be satisfied).

[0265]

[0266] In the formula: and are the slack amounts of the upper and lower limits of the adaptation range of the power commands; S i,max is the maximum upward / downward adjustable amount of the power source determined by the adjustment speed and adjustment capacity. According to the net load prediction result, the direction of the system adjustment requirement can be judged, and then it is determined whether S i,max takes the maximum upward adjustment amount or the maximum downward adjustment amount; r i is the remaining adjustment margin of the power source outside the adaptation range. As can be seen from Equation (9), there are only two cases where the adaptation range of the power commands is greater than or equal to a certain value or less than or equal to a certain value. Therefore, for each power source, the remaining adjustment margin of the power source outside the adaptation range is a unique fixed value; M is a constant with a very large value.

[0267] Equation (11) means that the constraint of the adaptation range of the power commands is not forced to be satisfied. When it conflicts with other operation constraints, the AGC commands allocated to the power source can exceed its adaptation range; Equations (12) and (13) mean that the AGC commands received by the power source after relaxation should be within its adjustable range; the penalty for the slack amount of the adaptation range of the power commands is added in Equation (14), so that the AGC commands exceeding its adaptation range are only allocated to the power source when the system adjustment requirements cannot be met. In addition, a coefficient 1 / r related to the remaining adjustment margin of the power source outside the adaptation range is introduced into the penalty term i , so as to make as few power sources as possible exceed their command adaptation ranges to ensure the response performance of the power sources in the system to the received AGC commands.

[0268] To meet the requirements of computational efficiency, the current AGC real-time allocation strategy generally adopts preset rules to quickly allocate instructions. The basic idea of the allocation rule proposed in the present invention is to allocate in the order of regulation speed. Meanwhile, on the premise of meeting the system regulation requirements, instructions within the adaptation range of each power source are allocated to ensure the response performance of the power source to the received AGC instructions. The process of the proposed AGC real-time allocation strategy is as Figure 3 shown (taking the allocation of the upward frequency modulation AGC instruction as an example). The proposed allocation strategy is not limited to the basic rule of allocating in the order of regulation speed. For other common rules adopted in AGC real-time allocation, the adaptation range of the power source instructions can also be embedded in a corresponding manner.

[0269] The specific allocation steps are described in detail as follows:

[0270] 1) Allocate the total system AGC instruction (A sys ) to each power source in the order of decreasing regulation speed. The upper limit of the AGC instruction received by each power source is the upper limit (A i,max ) of its adaptation range, and the lower limit is 0.

[0271] 2) Determine whether there are still unallocated instructions (A sys,re ) in the system. If there are unallocated instructions, it indicates that only considering the adaptation range of the power source instructions cannot fully meet the system regulation requirements, and go to step 3); otherwise, go to step 4).

[0272] 3) Considering the actual regulation upper limit of the power source determined by the regulation speed and regulation capacity, calculate the remaining adjustable amount of each power source, and allocate the remaining instructions (A sys,re ) to each power source in the order of decreasing remaining adjustable amount, which can reduce the number of power sources whose AGC instructions exceed the upper limit of their own adaptation range, and go to step 10) to complete this instruction allocation.

[0273] 4) Determine whether the instruction (A j ) received by the last power source receiving the instruction (denoted by j) is within its adaptation range [A j,min , A j,max . If it is within the adaptation range, it indicates that all power sources can complete this regulation task within the instruction adaptation range, and go to step 10) to complete this instruction allocation; otherwise, combined with the judgment in step 2), it indicates that the AGC instruction received by this power source is less than the lower limit (A j,min ) of its adaptation range, and go to step 5).

[0274] 5) Determine whether the power supply j of the last received instruction is the first power supply with the fastest adjustment speed. If not, it indicates that at least two power supplies are involved in this adjustment and the instruction received by at least the first power supply is within its adaptation range. Go to step 6); otherwise, it indicates that the system adjustment requirement is small, only the first power supply receives the instruction and the instruction does not reach the lower limit of the power supply adaptation range. At this time, the power supplies receiving the instruction are not within their adaptation ranges. Go to step 7).

[0275] 6) Calculate the difference (denoted as R j ) between the instruction (A j,min ) currently received by the power supply j and the lower limit (A j ) of its adaptation range, indicating that the instruction received by the power supply j needs to be increased by R j to make it meet the instruction adaptation range. Further, calculate the total amount of adjustment that can be reduced by other power supplies receiving instructions under the condition of meeting their own adaptation ranges, as shown in Equation (15).

[0276]

[0277] In the formula: D represents the total amount of adjustment that can be reduced by other power supplies receiving instructions under the condition of meeting their own adaptation ranges (since the method of sequential distribution according to the adjustment speed is adopted in step 1), so for other power supplies receiving instructions, the instruction they receive is the upper limit of their adaptation range); I A represents the set of power supplies receiving instructions.

[0278] Subsequently, judge the size of R j and D. If R j is greater than D, it indicates that the instruction of the power supply j cannot be adjusted to its adaptation range under the condition that other power supplies all meet their own instruction adaptation ranges. Do not adjust the current instruction distribution result. Go to step 10) to complete this instruction distribution; otherwise, reduce the instructions received by other power supplies in the order from slowest to fastest adjustment speed (let the power supply with a faster adjustment speed undertake more adjustment tasks). The reduced instructions must not exceed the power supply's own adaptation range, and increase the instructions received by the power supply j equally until it reaches the lower limit (A j,min ) of the adaptation range. Go to step 10) to complete this instruction distribution.

[0279] 7) Judge whether the overall system adjustment requirement is less than the minimum value of the lower limits of all power supply adaptation ranges. If so, it indicates that the system adjustment requirement is too small to perform instruction distribution under the condition that some power supplies meet their adaptation ranges. Do not re - distribute the AGC instructions. Go to step 10) to complete this instruction distribution; otherwise, it indicates that the system adjustment requirement exceeds the lower limits of the adaptation ranges of some power supplies, and the system can perform instruction distribution under the condition that some power supplies meet their adaptation ranges. Go to step 8) to re - distribute the AGC instructions.

[0280] 8) Select power supplies whose lower limit of the adaptation range is less than the system regulation requirement. Taking the upper limit of the power supply adaptation range as the limit value, re - allocate AGC commands to these power supplies in descending order of the regulation speed.

[0281] 9) Determine whether there are still un - allocated commands in the system. If there are, it indicates that only considering the power supply command adaptation range cannot fully meet the system regulation requirement. At this time, allocate the remaining commands to other power supplies that have not received commands in descending order of the regulation speed, and go to step 10) to complete this command allocation; otherwise, return to step 4), and adjust the AGC command allocation result within the selected power supply range to make the AGC commands received by each power supply within its adaptation range as much as possible. It should be noted that at this time, j must be greater than 1, and the process will not fall into an infinite loop.

[0282] 10) Obtain the command allocation results of all power supplies and send them to each power supply.

[0283] The case simulation based on the actual data of gas - fired units and hydro - power units in a provincial power grid of China shows that even using traditional data - driven technologies can achieve relatively accurate perception of power supply frequency - modulation performance (the average perception accuracy reaches 94.36% and the average perception precision reaches 100%), verifying the effectiveness and feasibility of the proposed framework for extracting the power supply command adaptation range. Other more advanced data - driven technologies can also be used in the framework for extracting the power supply command adaptation range proposed in the present invention, which is expected to further improve the accuracy of power supply frequency - modulation performance perception and the extraction effect of the command adaptation range.

[0284] Example 13:

[0285] Verification of the AGC regulation demand decomposition method for dynamically matching the real - time operating conditions of power supplies, the steps include:

[0286] Construct a power supply frequency - modulation performance perception model

[0287] Taking the historical multi - dimensional operating conditions of the power supply and the received historical AGC commands as sample inputs, and taking whether the power supply historical frequency - modulation performance meets the standard at the corresponding moment as the sample output, construct a frequency - modulation performance compliance perception model based on SVM.

[0288] Extract the power supply command adaptation range

[0289] Based on the power supply frequency - modulation performance compliance rules obtained by the SVM model in (1), reverse - infer the command range that can make its frequency - modulation performance meet the standard according to the current operating conditions of the power supply. Subsequently, combined with operating parameters such as the power supply regulation speed and regulation capacity, obtain the final power supply command adaptation range.

[0290] AGC advanced optimization allocation with embedded power supply command adaptation range

[0291] According to the instruction adaptation range in (2), it is embedded in the AGC lead optimization distribution strategy to achieve the dynamic matching of AGC instruction issuance and power operation conditions.

[0292] AGC real-time distribution with embedded power instruction adaptation range

[0293] According to the instruction adaptation range in (2), it is embedded in the AGC real-time distribution strategy, and is distributed in the order of adjustment speed. At the same time, on the premise of meeting the system adjustment requirements, instructions that conform to their adaptation ranges are distributed to each power source to ensure the response performance of the power source to the received AGC instructions.

[0294] The specific simulation results are as follows:

[0295] 1) Example description

[0296] The comparison methods in the simulation are as follows:

[0297] M1: A power frequency modulation performance perception model based on SVM, but without considering PCA and false alarm penalty factors;

[0298] M2: The same as M1, but further uses PCA to extract key features of the input features;

[0299] M3: The same as M2, but further embeds a false alarm penalty factor and uses a traditional iterative adjustment method to determine the specific value of the false alarm penalty factor;

[0300] M4 (the method proposed in the present invention): The same as M3, but uses the false alarm penalty factor iterative adjustment method proposed in the present invention.

[0301] The present invention collects the actual response data (including unit operation conditions, received AGC instructions, and unit frequency modulation performance indicators) of 2 gas turbines (characterized as T1, T2) and 1 hydropower plant (including a total of 5 hydropower units, characterized as H1) in a provincial power grid of our country to verify the effectiveness of the proposed method. The 2 gas turbines collected by the present invention receive grid AGC instructions respectively during actual operation, belonging to Mode 1, so they are used as 2 regulating power sources in the simulation; the hydropower plant receives grid AGC instructions as a whole during actual operation, and the frequency modulation performance indicators also reflect the overall performance of the power plant, belonging to Mode 2, so the 5 hydropower units are used as a whole as a regulating power source in the simulation. In the simulation, the present invention establishes a frequency modulation performance perception model for 3 regulating power sources respectively.

[0302] In addition, it is worth emphasizing that the present invention conducts case studies based on actual power supply data. During the case study process, due to considerations of power grid operation safety, it is difficult for the present invention to actually issue AGC commands to the power supply that match the extracted adaptation range, and verify the effectiveness of the command adaptation range extraction method and AGC allocation strategy by observing whether the frequency regulation performance of the power supply meets the standards. In this regard, the present invention will first verify the accuracy of the power supply frequency regulation performance perception model. Since the constructed frequency regulation performance perception model can establish a linear analytical relationship among the power supply operation conditions, the compliance of the power supply frequency regulation performance, and the AGC commands, under the condition of accurate power supply frequency regulation performance perception, the extracted power supply command adaptation range will be accurate. Subsequently, the effectiveness of the AGC allocation strategy can be verified by comparing whether the commands received by the power supply are within its adaptation range under different AGC allocation strategies.

[0303] Next, the effectiveness of the power supply frequency regulation performance perception model based on SVM and the AGC allocation strategy with the embedded power supply command adaptation range proposed by the present invention will be verified respectively.

[0304] The computational performances of the four power supply frequency regulation performance perception models M1 - M4 are shown in Table 1. Among them, the accuracy and precision are the average statistical results of 3 power supply frequency regulation performance perception models. In addition, the present invention constructs corresponding frequency regulation performance perception models for up - frequency regulation and down - frequency regulation respectively, and the data listed in Table 1 are the comprehensive statistical results of the two types of models (the training time consumption is the average training time of the two types of models).

[0305] Table 1 Computational Performances of M1 - M4

[0306]

[0307] By comparing M1 and M2, it can be found that after PCA processing, M2 can more effectively explore the influence of key input information on the unit frequency regulation performance, avoid the interference of redundant information, and thus effectively improve the accuracy and precision of frequency regulation performance perception.

[0308] By comparing M2 and M3, it can be found that M2 does not specifically handle the false alarm phenomenon, resulting in a low precision rate of M2. Based on this, instructions that do not actually match the power operation conditions are more likely to appear in the extracted adaptation range, thus affecting the effectiveness of subsequent AGC instruction allocation. In contrast, M3 embeds a false alarm penalty factor, which can effectively improve the precision rate of the perception model (the perception precision rate of the test set reaches 100% in this example), reduce the false alarm phenomenon, thereby alleviating the situation where inadaptable instructions are included in the adaptation range, and improving the effectiveness of the instruction adaptation range extraction result to a certain extent. However, the focus on false alarm samples leads to a certain degree of weakening of the penalty for missed alarm samples in the frequency modulation performance perception model, resulting in a slight decrease in the accuracy rate of the perception model. However, as described in Section 2.2, the improvement of the precision rate is more important for the power instruction adaptation range extraction method described in the present invention.

[0309] Similar to M3, M4 also embeds a false alarm penalty factor, and the values of the false alarm penalty factor are similar after iterative adjustment. Therefore, the perception accuracy rate and precision rate of M3 and M4 are finally the same. The main difference between the two lies in the iterative adjustment method of the false alarm penalty factor. The method M4 proposed in the present invention can more efficiently find the most suitable false alarm penalty factor through the adjustment of the iterative step size of the false alarm penalty factor, and the training efficiency is increased by 8.83 times compared with M3. For the AGC instruction allocation with a short calculation period (generally in seconds), the proposed method can reserve more time for the subsequent extraction of the instruction adaptation range and the allocation of AGC instructions.

[0310] To further illustrate the effectiveness of the false alarm penalty factor adjustment method proposed in the present invention, taking the down-frequency modulation performance perception model of the gas turbine unit T1 as an example, the iterative adjustment processes of the false alarm penalty factors of M3 and M4 are as Figure 4 shown (PA and PP in the figure are the accuracy rate and precision rate of the training set, which are used to guide the adjustment of the false alarm penalty factor). The final false alarm penalty factors of M3 and M4 are 4.37 and 4.40 respectively.

[0311] From Figure 4It can be seen that the proposed method M4 can quickly locate the penalty factor that enables the classification accuracy to reach 100% by adopting a larger adjustment step size. This penalty factor will be slightly larger than the penalty factor that can just make the classification accuracy reach 100% found by a finer step size. In this example, since the iterative adjustment process of increasing the penalty factor for the last time to make the classification accuracy reach 100% does not cause a decrease in the classification accuracy, the proposed method does not fine-tune the penalty factor. While M3 adopts the traditional adjustment method and searches for the penalty factor that can just make the classification accuracy reach 100% through a finer step size, and the excessive number of adjustment times results in low training efficiency. Thus, for the down-regulation frequency performance perception model in this example, M4 can effectively reduce the iterative adjustment times of the false alarm penalty factor, and finally improve the training efficiency by 10.97 times under the condition that the training effect is the same as that of M3.

[0312] In summary, the power frequency modulation performance perception model based on SVM proposed in the present invention can accurately predict the frequency modulation performance of the power supply under given operating conditions and AGC commands, indicating that relatively accurate power frequency modulation performance perception can be achieved even by using conventional data-driven technologies, verifying the effectiveness and feasibility of the proposed power command adaptation range extraction framework.

[0313] 2) Verification of the effectiveness of the AGC allocation strategy with the embedded power command adaptation range

[0314] ① In the simulation, the parameters of the 3 regulating power supplies and the extracted command adaptation ranges are shown in Table 2. Among them, the upper limit of the adjustable capacity is the maximum allocation command set when the power grid allocates AGC commands to the corresponding power supply.

[0315] Table 2 Power regulation parameters and command adaptation ranges

[0316]

[0317] The comparison methods in the simulation are as follows:

[0318] M5: The conventional AGC command lead allocation strategy without considering the power command adaptation range;

[0319] M6 (the method proposed in the present invention): On the basis of M5, further consider the power command adaptation range;

[0320] M7: The conventional AGC real-time allocation strategy based on the order of power regulation speed;

[0321] M8 (the method proposed in the present invention): On the basis of M7, further consider the power command adaptation range.

[0322] The effectiveness of the AGC allocation strategy will be verified by comparing whether the commands received by the power supply are within its adaptation range under different AGC allocation strategies.

[0323] ②Taking frequency modulation as an example, compare the AGC command lead optimization allocation results of two methods, M5 and M6, under different system regulation requirements. The allocation results of the two allocation strategies, M5 and M6, under different regulation requirements are shown in Table 3. In the table, Y and N in parentheses indicate whether the commands received by the units are within or outside the adaptation range.

[0324] Table 3 AGC command allocation results of M5 and M6 under different regulation requirements

[0325]

[0326] As can be seen from Table 3, since the objective function of the lead optimization allocation is to minimize the regulation cost, under different system regulation requirements, M5 will sequentially call the full adjustable capacity of each power source in ascending order of regulation cost. The regulation cost of this allocation method is relatively small, but since the command adaptation range of the power source is not considered, the commands issued by M5 are difficult to match the operating conditions of the power source, and the command response effect of the power source is difficult to guarantee.

[0327] In contrast, the method M6 proposed in the present invention takes into account the command adaptation range of the power source, and the issued commands can better fit the command adaptation range of the power source: for the scenario where the system regulation requirement is 27 MW, the sum of the upper limits of the adaptation ranges of the three power sources can meet the total regulation requirement, so in M6, the commands received by all units are within their adaptation ranges. Compared with M6, in M5, the command received by power source H1 exceeds its adaptation range; for the scenario where the system regulation requirement is 99 MW, since the sum of the upper limits of the adaptation ranges of the three power sources cannot meet the total system regulation requirement, some power sources need to exceed the command adaptation range. Considering that the remaining regulation margin of power source H1 outside the adaptation range is the largest, the remaining regulation requirement is allocated to H1, so that as few power sources as possible exceed their command adaptation ranges to ensure the response performance of the power sources in the system to the received commands. Compared with M6, in M5, the commands received by power sources T2 and H1 do not conform to their adaptation ranges.

[0328] In summary, the AGC commands allocated by the proposed method will better fit the command adaptation range of the power source, thus ensuring the command response performance of the power source and helping to improve the system frequency quality. Although the regulation cost of the proposed method is higher than that of the conventional method with minimizing the regulation cost as the primary objective, which affects the system regulation economy, for AGC command allocation, ensuring the command response performance of the power source and improving the system frequency quality are of more important significance.

[0329] ③Verification of the effectiveness of the AGC real-time allocation strategy with embedded power source command adaptation range

[0330] Taking frequency regulation as an example, the real-time AGC command allocation results of two methods, M7 and M8, under different system regulation requirements are compared. As can be seen from Table 2, in this example, the upper limit of the command adaptation range of each power source is the same as the upper limit of its own adjustable capacity. At this time, the proposed method M8 can be regarded as a readjustment of the command allocation result of the conventional method M7.

[0331] Under different regulation requirements, the allocation results of the two allocation strategies, M7 and M8, are shown in Table 4. Since both the conventional method M7 and the proposed method M8 are based on the regulation speed sequence as the basic basis and do not pay attention to the regulation economy, the system regulation cost is not shown in Table 4.

[0332] Table 4 AGC command allocation results of M7 and M8 under different up-regulation requirements

[0333]

[0334] As can be seen from Table 4, for the scenario where the system regulation requirement is 118 MW, under the conventional allocation method M7 according to the regulation speed sequence, the commands received by all power sources are within their adaptation ranges. Therefore, the proposed method M8 does not adjust the command allocation result, and at this time, the allocation results of M7 and M8 are the same. For the scenario where the system regulation requirement is 112 MW, under the conventional method M7, the power sources H1 and T1 with faster regulation speeds will be called first, resulting in the command received by the power source T2 being too small and not within the command adaptation range. In response to this situation, the proposed method M8 adjusts the command allocation result by reducing the regulation amount of the power source T1 and synchronously increasing the regulation amount of the power source T2, so that the commands received by all power sources are within their adaptation ranges. For the scenario where the system regulation requirement is 6 MW, under the conventional method M7, the power source H1 with the fastest regulation speed will undertake all the regulation tasks. However, due to the too small system regulation requirement, the command received by H1 at this time is not within its adaptation range. In response to this situation, the proposed method M8 combines the lower limit of the command adaptation range of each power source to redistribute the system regulation requirement, and the power unit T1 will undertake all the regulation tasks, and the allocation result meets its adaptation range.

[0335] In summary, through the embedding of the power source command adaptation range, the proposed method M8 improves the matching degree between AGC command allocation and the power source operation state, which helps to improve the system frequency quality.

[0336] Aiming at the problem that the existing AGC allocation strategy does not consider the difference in the response effect of power commands under different operating conditions, the present invention proposes an AGC regulation demand decomposition strategy for dynamically matching the real-time operating conditions of power sources. First, a power frequency modulation performance perception model based on SVM is constructed. Based on this, the command range that the power source can effectively respond to under different operating conditions is extracted. At the same time, aiming at the false alarm situation of frequency modulation performance, an adaptive penalty factor is introduced to improve the accuracy of the perception model, thereby improving the extraction effect of the command adaptation range. On this basis, for the current commonly used AGC command lead allocation and real-time allocation, an AGC allocation strategy embedded with the power command adaptation range is proposed to improve the matching degree between the AGC command allocation result and the real-time operating conditions of the power source. The case simulation based on the actual power data of a provincial power grid in China shows that the proposed frequency modulation performance perception model can accurately predict the frequency modulation performance of the power source under the given operating conditions and AGC command conditions, which can strongly support the extraction of the power command adaptation range. The proposed AGC allocation strategy embedded with the power command adaptation range can allocate adjustment commands matching the operating conditions of each power source to each power source on the premise of meeting the system regulation requirements, thereby ensuring the power command response performance and helping to improve the system frequency quality.

Claims

1. A method for decomposing AGC regulation requirements that dynamically matches the real-time operating conditions of a power supply, characterized in that: The following steps are involved: 1) Build a power frequency modulation performance perception model. 2) Set the historical frequency regulation performance vector of the power supply as the standard parameter, and input the current multi-dimensional operating conditions of the power supply into the power supply frequency regulation performance perception model to extract the power supply AGC instruction adaptation range; 3) The power supply AGC instruction adaptation range is embedded in the AGC allocation strategy to achieve dynamic matching between the AGC instruction issuance and the power supply operating conditions.

2. The AGC regulation demand decomposition method for dynamically matching the real-time operating conditions of the power supply according to claim 1 is characterized in that: In step 1), the power frequency modulation performance perception model includes a first power frequency modulation performance perception model and a second power frequency modulation performance perception model; The input of the first power frequency regulation performance perception model is the power multi-dimensional operating condition and the upper frequency regulation AGC instruction, and the output is the power frequency regulation performance vector at the corresponding moment; The input of the first power frequency regulation performance perception model is the power multi-dimensional operating condition and the frequency regulation AGC instruction, and the output is the power frequency regulation performance vector at the corresponding moment; The operating conditions include the operating conditions of gas units and hydropower units; The gas unit operating conditions include the unit's current output, gas flow control instructions, air compressor inlet temperature, and main fuel flow control valve position; The operating conditions of the hydropower unit include the current output of the unit, water head, and guide vane opening.

3. The AGC regulation demand decomposition method for dynamically matching the real-time operating conditions of the power supply according to claim 2 is characterized in that: The steps of constructing the first power frequency modulation performance perception model include: 1) Obtain the historical multi-dimensional operating conditions of the power supply, the received historical frequency modulation AGC instructions, and the historical frequency modulation performance vector of the power supply at the corresponding moment; the frequency modulation performance vector includes a standard parameter or a non-standard parameter; the standard parameter is 1, and the non-standard parameter is 0; 2) Construct a sample data set with the historical multi-dimensional operating conditions of the power supply and the received historical frequency modulation AGC instructions as input, and the historical frequency modulation performance compliance results of the power supply at the corresponding moment as output; 3) Using the sample data set to train and verify the SVM binary classification model, a first power supply frequency modulation performance perception model is obtained. The decision function f(x PCA ) is as follows: f(x PCA )=w T x PCA +b (1) Where: w and b represent the normal vector and intercept of the SVM decision hyperplane respectively; x PCA Represents input.

4. The AGC regulation demand decomposition method for dynamically matching the real-time operating conditions of the power supply according to claim 2 is characterized in that: The steps of constructing the second power supply frequency modulation performance perception model include: 1) Obtain the historical multi-dimensional operating conditions of the power supply, the received historical frequency modulation AGC instructions, and the historical frequency modulation performance vector of the power supply at the corresponding moment; the frequency modulation performance vector includes a standard parameter or a non-standard parameter; the standard parameter is 1, and the non-standard parameter is 0; 2) Construct a sample data set with the historical multi-dimensional operating conditions of the power supply and the received historical down-frequency AGC instructions as input, and the historical frequency regulation performance of the power supply at the corresponding moment as output; 3) Use the sample data set to train and verify the SVM binary classification model to obtain the second power supply frequency modulation performance perception model.

5. The AGC regulation demand decomposition method for dynamically matching the real-time operating conditions of the power supply according to claim 2 is characterized in that: When constructing the sample data set, the PCA method is used to extract key features from the historical multi-dimensional operating conditions of the power supply and the historical AGC instructions received, and the extracted principal component features are used as input; The main component characteristics after repeated PCA processing are as follows: Where: k and K are the number and set of the retained principal components respectively; x PCA,k represents the kth principal component; n and N are the dimension number and set of the original input features respectively; c k,n x is the contribution rate of the original input feature of the nth dimension to the feature of the kth principal component; n are the features before processing.

6. The AGC regulation demand decomposition method for dynamically matching the real-time operating conditions of a power supply according to claim 3 or 4, characterized in that: During model training, the update targets of the normal vector and intercept w and b are as follows: Where: l and L are the number and set of training samples respectively; ||w|| is the norm of w, expressed as γ l is a slack variable indicating the degree of sample misclassification; C + and C - are the misclassification penalty factors for samples that meet the FM performance standards and samples that do not meet the standards; L + and L - are the sets of samples that meet the frequency modulation performance standards and samples that do not meet the standards. l Indicates the classification result.

7. The AGC regulation demand decomposition method for dynamically matching the real-time operating conditions of a power supply according to claim 3 or 4, characterized in that: During model training, the update process of the normal vector and intercept w and b is as follows: 1) Let the penalty factor C +,q =1, C -,q =1, and set the maximum number of iterations q max , initial iteration step length of false alarm penalty factor ΔC - , and the initial mark f = 0 representing the adjustment method of the false alarm penalty factor; 2) Use the sample data set to train the SVM and calculate the classification accuracy of this iteration based on the training set PP And classification accuracy PA ,Right now: Where TP and FP are the number of true positive examples and false positive examples; TN and FN are the number of true negative examples and false negative examples; 3) If the number of iterations has reached the maximum number of iterations q max , then go to step 8), otherwise, go to step 4); 4) If f=0, an adjustment method of increasing the false alarm penalty factor is adopted, and the process goes to step 5); otherwise, an adjustment method of reducing the false alarm penalty factor is adopted, and the process goes to step 7); 5) If It indicates that there is no false positive in the training set, go to step 6), otherwise, let C -,q+1 =C -,q +ΔC - , q=q+1, return to step 2); 6) If This indicates that the increase in the false positive penalty factor in this iteration does not cause a decrease in classification accuracy, so go to step 8). Otherwise, this indicates that the increase in the false positive penalty factor in this iteration is too large, so let C -,q+1 =C -,q -ΔC - / 10, q=q+1, return to step 2); 7) If Let C -,q+1 =C -,q -ΔC - / 10, q=q+1, return to step 2); otherwise, it indicates that continuing to reduce the false positive penalty factor will affect the classification accuracy, and the q-1th iteration is taken as the final result, and go to step 8); 8) Obtain the optimal parameter w of the SVM model * and b * , output the result.

8. The method for decomposing the AGC regulation requirements according to claim 1, characterized in that: The power supply AGC command adaptation range is as follows: Where: a is the input feature dimension number corresponding to the AGC instruction; x a AGC command received by the power supply; c k,a is the contribution rate.

9. The method for decomposing the AGC regulation requirements according to claim 1 is characterized in that: The objective function and constraints of the AGC allocation strategy are as follows: Where: i is the power supply number; I is the power supply set; b A,i is the power conditioning cost; A i is the power supply regulation amount, that is, the AGC instruction received by the power supply; and S is the relaxation amount of the upper and lower limits of the power command adaptation range; i,max is the maximum upward / downward adjustable amount of the power supply determined by the adjustment speed and adjustment capacity; M is a constant with a very large value; A i,max , A i,min It is the upper and lower limits of the power supply regulation.

10. The method for decomposing the AGC regulation requirements according to claim 1, characterized in that: The steps to achieve dynamic matching between AGC instruction issuance and power supply operating conditions include: 1) Adjust the system total AGC instruction A in order from fast to slow. sys Distributed to each power supply; the upper limit of the AGC command received by each power supply is the upper limit of its adaptation range A i,max , the lower limit is 0; 2) Determine whether the system has any unassigned instructions A sys,re If there are unassigned commands, it means that only considering the power command adaptation range cannot fully meet the system regulation requirements, go to step 3); otherwise, go to step 4). 3) Considering the actual upper limit of power supply regulation determined by regulation speed and regulation capacity, calculate the remaining adjustable amount of each power supply, and sort the remaining instructions A in descending order of the remaining adjustable amount. sys,re Distribute to each power supply, and go to step 10) to complete this instruction distribution. 4) Determine the instruction A received by the last power supply j that receives the instruction j Is it within its adaptation range [A j,min ,A j,max ], if it is within the adaptation range, it means that all power supplies can complete the adjustment task within the instruction adaptation range, and go to step 10) to complete the instruction allocation; otherwise, go to step 5). 5) Determine whether the last power supply j to receive the instruction is the first power supply with the fastest adjustment speed. If not, go to step 6); otherwise, go to step 7). 6) Calculate the instruction A currently received by power supply j j And the lower limit of the adaptation range A j,min The difference R j , and calculate the total amount of adjustment D that can be reduced by other power supplies that receive the command under the condition that they meet their own adaptation range, and then determine R j and the size of D, if R j If it is greater than D, go to step 10) to complete the instruction allocation. Otherwise, reduce the instructions received by other power supplies in the order of adjustment speed from slow to fast. The reduced instructions must not exceed the adaptation range of the power supply itself, and increase the instructions received by power supply j by the same amount until it reaches the lower limit A of the adaptation range. j,min , go to step 10) to complete the instruction allocation; The total amount of adjustment D that can be reduced is as follows: 7) Determine whether the overall regulation demand of the system is less than the minimum value of the lower limit of all power adaptation ranges. If so, do not reallocate the AGC instruction, and go to step 10) to complete the instruction allocation; otherwise, go to step 8) to reallocate the AGC instruction; 8) Filter out the power supplies whose lower limit of the adaptation range is less than the system regulation requirement, and use the upper limit of the power adaptation range as the limit value to reallocate AGC instructions to these power supplies according to the regulation speed from fast to slow; 9) Determine whether the system has any unassigned instructions. If so, assign the remaining instructions to other power supplies that have not received instructions in the order of adjustment speed from fast to slow, and go to step 10) to complete the instruction assignment; otherwise, return to step 4) and adjust the AGC instruction assignment result within the screened power supply range; 10) Obtain the command allocation results of all power supplies and send them to each power supply.

Images