Method for extracting scheduling rules of multi-energy complementary power intelligent system

Abstract

The application provides a multi-energy complementary power intelligent system scheduling rule extraction method, comprising the following steps: constructing a past external condition vector sequence, performing a simulation trace accumulation process, generating and extracting stable landscape features from a trace tensor, constructing an initial rule template library, collecting real-time operation data, constructing a real-time external condition vector, obtaining a real ecological niche offset and an ecological niche construction potential of each energy unit, and updating an online trace concentration matrix, generating an online trace vector based on the online trace concentration matrix, and matching the online trace vector with a base vector of the stable landscape features, outputting an optimal matching landscape feature or a no matching flag, retrieving a rule template based on the optimal matching landscape feature, obtaining a final value of a correction amount of each energy unit, performing a safety constraint check on a tentative output instruction, issuing a formal scheduling instruction after the check is passed, and dynamically updating a confidence degree of a corresponding rule template combined with a comprehensive feedback factor, so as to ensure the accuracy and adaptability of the extracted scheduling rule.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and more specifically, to a method for extracting scheduling rules for multi-energy complementary intelligent power systems. Background Technology

[0002] With the continuous increase in the proportion of new energy installed capacity in the power system, multi-energy complementary systems have become an important way to absorb clean energy and ensure power supply reliability. Dispatch rule extraction is an important technical link of multi-energy complementary power intelligent system. It mines typical dispatch patterns from historical operation data and generates reusable dispatch decision rules. However, traditional dispatch rule extraction methods have certain limitations.

[0003] On the one hand, traditional rule extraction relies on manual experience tuning and mechanism model derivation, which can only form fixed threshold rules covering a few typical operating conditions. This often makes it difficult to adapt to the complex and ever-changing operating conditions under high-proportion new energy scenarios. Moreover, the quality of the rules is highly dependent on the professional experience of the personnel, and most of them lack the mining of the nonlinear and multivariate coupled scheduling behavior patterns hidden in historical operating data, resulting in insufficient rule coverage and refinement.

[0004] On the other hand, existing data-driven rule extraction methods often lack the extraction and mining of the temporal cumulative effect of scheduling behavior and the collaborative characteristics of multiple units, resulting in limited applicability and control accuracy of the rules. Summary of the Invention

[0005] In view of the aforementioned problems of insufficient rule coverage and refinement, and limited applicability and control precision, in conjunction with the first aspect of the present invention, embodiments of the present invention provide a method for extracting scheduling rules for multi-energy complementary power intelligent systems, the method comprising: S1: Based on past system operation data, construct past external condition vector sequences, establish basic ecological niche functions and generate past offset sequences, define simulated trace accumulation process and generate trace tensors and initial trace matrices, extract stable geomorphic features from trace tensors, and simultaneously construct an initial rule template library. S2: During the online operation phase, real-time operation data is collected according to the preset control cycle, and a real-time external condition vector is constructed. After the preset confidence gating judgment, the actual niche offset and niche construction potential of each energy unit are obtained, and the online trace concentration matrix and niche memory matrix are updated synchronously. S3: Generate online trace vectors based on the online trace concentration matrix, and match them with the basis vectors corresponding to each group of stable landform features, outputting the best matching landform feature or no matching flag; S4: When a match is successful, the corresponding rule template is retrieved from the initial rule template library based on the optimal matching geomorphic features. Simultaneously, the current working condition base value, matching similarity and niche construction potential are combined to obtain the absolute output correction amount of each energy unit. After being limited by the basic niche boundary, the final value of the correction amount is output. S5: Perform safety constraint verification on the tentative output command after superimposed correction final value. After the verification is passed, issue formal scheduling command, collect actual operation data after command execution, and dynamically update the confidence level of the corresponding rule template in combination with comprehensive feedback factor.

[0006] Compared with the prior art, the present invention has the following beneficial effects: Step S1 automatically completes the construction of basic ecological niche functions, the generation of trace tensors and the extraction of geomorphic features based on historical operation data. It does not rely on manual experience to adjust rule thresholds. Through tensor decomposition, it mines the implicit scheduling patterns of multi-unit and multi-trace dimension coupling, and generates rule templates covering multiple typical working conditions, thereby effectively capturing nonlinear scheduling patterns in historical data and expanding the rule coverage and refinement. Step S2 completes the quantification of the actual ecological niche shift, the acquisition of the ecological niche building potential, and the dynamic update of the online trace concentration by collecting operational data in real time, thereby providing comprehensive and accurate data support for subsequent steps. Step S3 uses online trace vectors to match the basis vectors corresponding to each group of stable landform features to identify the current scheduling mode, thereby achieving the matching of the optimal typical mode in the multi-dimensional trace space and adapting to complex operating conditions under different load levels and new energy penetration rates, thus improving the accuracy of scheduling mode identification. Step S4 calls the corresponding rule template based on the matching result, dynamically obtains the absolute output correction amount by combining the matching similarity, working condition base value and niche construction potential, and combines the basic niche boundary limiting mechanism to achieve adaptive fine-tuning of the called rules, avoiding the rigidity of using fixed threshold rules, thereby achieving adaptation to different operating states under the same working conditions. Step S5 ensures the operational security of rule implementation through multi-level security constraint verification, and obtains comprehensive feedback factors based on actual operational effects to dynamically adjust rule confidence, thereby achieving the selection and elimination of inferior rules in the rule template library and ensuring the adaptability of scheduling rules. Attached Figure Description

[0007] Figure 1 This is a flowchart of the steps in the method for extracting scheduling rules for a multi-energy complementary intelligent power system according to the present invention; Figure 2 This is a flowchart illustrating step S1 in the multi-energy complementary power intelligent system scheduling rule extraction method of the present invention. Detailed Implementation

[0008] The present invention will now be described in detail with reference to the accompanying drawings. Figure 1 This is a flowchart illustrating the steps of the multi-energy complementary power intelligent system scheduling rule extraction method of the present invention. Figure 2 This is a flowchart of step S1 in the method for extracting scheduling rules for a multi-energy complementary intelligent power system according to the present invention. The following is a detailed introduction to the method for extracting scheduling rules for a multi-energy complementary intelligent power system.

[0009] Step S1: Based on past system operation data, construct a sequence of past external condition vectors, establish a basic ecological niche function and generate a sequence of past offsets, define the simulated trace accumulation process and generate a trace tensor and an initial trace matrix, extract stable landform features from the trace tensor, and simultaneously construct an initial rule template library.

[0010] Specifically, step S1 includes steps S11-S14: Step S11: Based on past system operation data, construct a sequence of past external condition vectors. Specifically: First, obtain historical system operation data covering the entire operating cycle. This historical system operation data must include at least the following: Total system load ; Actual output value of each energy unit subscript Corresponding to each energy unit in a multi-energy complementary power intelligent system; The ultra-short-term power forecasts for new energy sources should include at least wind power forecasts. Compared with photovoltaic power forecast Each predicted data point is accompanied by a corresponding prediction confidence level. and ; The state of charge value of an energy storage unit is only valid for that energy storage unit. The over-limit status records for each transmission line are as follows: a value of 0 indicates normal power flow, and a value of 1 indicates an over-limit has occurred. The ramp-up event record for each energy unit is the absolute value of the change in the actual output value of each energy unit in the current period relative to the actual output value in the previous period, divided by the rated capacity of that unit.

[0011] Next, a multidimensional external condition vector is constructed for each historical period. The multidimensional external condition vectors for all periods together constitute a sequence of past external condition vectors. The multidimensional external condition vectors contain at least the following dimensions: System load factor, which is the ratio of the total system load to the rated total system load, is used to characterize the current load level of the system. The higher the value, the greater the system load pressure. The renewable energy penetration rate, which is the ratio of the sum of the predicted wind power and the predicted photovoltaic power to the total system load, is used to characterize the proportion of renewable energy output in the system load. The higher the value, the stronger the system's dependence on renewable energy and the higher the scheduling uncertainty. Time period labels are scalar values ​​for each time period within a day that are mapped to a preset continuous numerical range. The seasonal coefficient is the division of the year into several segments according to seasons or months, and assigning different discrete or continuous values.

[0012] It should be noted that, due to the significant differences in the physical meaning and numerical range of each dimension, directly using them for distance calculation will lead to the large-value dimension dominating the calculation results. Therefore, it is necessary to perform Z-score standardization on the external condition vectors of all past periods dimension by dimension, and finally obtain the standardized sequence of past external condition vectors. .

[0013] In one possible embodiment, a multi-energy complementary system in a certain region is used as the application scenario. The system is set to include two conventional thermal power units, one centralized wind farm, one photovoltaic power station, and one energy storage system, totaling four energy units participating in dispatch. The system includes six main transmission lines. The rated total load of the system is 200MW, the rated capacity of a single thermal power unit is 120MW and 80MW respectively, the rated capacity of the wind farm is 50MW, the rated capacity of the photovoltaic power station is 40MW, and the rated power of the energy storage is 30MW. Suppose that during a certain historical period Total system load MW, the actual output of Unit 1 thermal power unit MW, actual output of Unit 2 thermal power unit MW, actual wind power output is 25MW, actual solar power output is 12MW, wind power power forecast value MW and corresponding confidence level Photovoltaic power forecast MW and corresponding confidence level All line over-limit statuses are 0. Assuming that the actual output of Unit 1 in the previous period was 88MW, its corresponding ramp amount is |90-88| / 120≈0.0167. Similarly, the ramp amount corresponding to each energy unit in the current system can be obtained. Assuming that analysis of the latest 90-day historical data yields a system load rate with a mean of 0.62 and a standard deviation of 0.18, and a renewable energy penetration rate with a mean of 0.31 and a standard deviation of 0.15, for the aforementioned historical period... If the system load rate is 0.75 and the new energy penetration rate is 0.253, then the first dimension of the standardized past external condition vector is (0.75-0.62) / 0.18≈0.72, the second dimension is (0.253-0.31) / 0.15≈-0.38, and so on, to obtain the complete past external condition vector (for the sake of simplifying the calculation, the above values ​​are all example assumptions, and the specific values ​​need to be determined based on the actual situation).

[0014] Step S12: Establish the basic niche function and generate a historical offset sequence. Specifically: First, for each energy unit A base value function is constructed based on all past system operation data. With bandwidth function The base function is used to characterize the central trend of the energy unit output under a given operating condition, while the bandwidth function is used to characterize the normal fluctuation range of the output under that operating condition. The method for generating the basic function is as follows: Given any standardized external condition vector In this case, the K periods with the smallest Euclidean distance to the vector are retrieved from all previous periods, where K is a preset positive integer. The actual output value of the corresponding energy unit in the K periods is extracted, and the median of the actual output value in the K periods is used as the output of the base function. Similarly, the standard deviation of the actual output value over K time periods is used as the output of the base function.

[0015] Based on the aforementioned base function and bandwidth function, the basic niche output range of energy unit i under any operating condition can be defined, i.e. This range represents the historical normal operating range and can be used to measure the degree to which the actual output value of an energy unit deviates from the historical normal level.

[0016] Next, for each historical period Each energy unit in Subtract the output value of the base function under the corresponding operating condition from the actual output value of each energy unit during the period to obtain the corresponding actual ecological niche offset. When the actual niche offset is positive, it means that the actual output of the energy unit is higher than the historical median output under similar operating conditions, corresponding to a niche expansion state; when it is negative, it means that the actual output of the energy unit is lower than the historical median level, corresponding to a niche contraction state. After calculating the actual niche offset for all historical periods, the actual niche offsets of energy units in all historical periods are combined to form a historical offset sequence.

[0017] Step S13, define the simulated trace accumulation process and generate the trace tensor and initial trace matrix, specifically: First, initialize a trace concentration matrix of a preset size. For example, initialize a zero matrix with a size of 4 rows and 3 columns, with all elements initialized to 0, as the initial trace concentration matrix. Iterate through all past periods, and in each past period, multiply all elements in the trace concentration matrix by a preset exponential decay factor. ,in A positive real decay coefficient pre-set for domain experts is used to determine the half-life length of trace memory.

[0018] Next, based on the writing rules corresponding to the preset simulated traces, the new writing amount corresponding to the preset simulated traces of each energy unit is obtained by combining the system operation data of the current time period, and a current writing matrix of the same size as the trace concentration matrix is ​​generated, and then it is superimposed on the trace concentration matrix element by element. The updated trace concentration matrix is ​​used as a slice of the trace tensor in the current time period and is fully saved.

[0019] Finally, after traversing all previous periods, the corresponding trace tensor is generated. At the same time, the final trace concentration matrix at the end of the trace accumulation simulation is used as the initial trace matrix.

[0020] It should be noted that the preset simulation traces include at least opportunity traces, risk traces, and cooperation traces, specifically: The opportunity trace is used to reflect the relative output margin of the unit from the upper limit of the basic ecological niche, that is, the space in which the unit can increase output without exceeding the upper limit of historical normal fluctuations. The larger the value, the more sufficient the output margin. The corresponding writing rule can be expressed as: ; in, Represented as the first Each energy unit in the current period The new write volume of opportunity traces is the quantitative value of new traces added within a single control cycle. It has no unit, and the subscript 1 is the fixed type identifier of opportunity traces, used to distinguish between risk traces and collaborative traces. Represented as the first Each energy unit in External condition vector of time period The base output value under the corresponding working condition, in MW, is calculated by the base value function in the basic niche function and is used to represent the historical concentration trend level of the unit's output under the corresponding working condition. Represented as the first Each energy unit in External condition vector of time period The output bandwidth under the corresponding operating condition, in MW, is calculated from the bandwidth function of the basic niche function and is used to characterize the historical normal fluctuation range of the unit's output under the corresponding operating condition. For the first Each energy unit in The actual operating output value for the time period, in MW, is the measurement data collected by the system in real time; At the same time, when the output value of the bandwidth function approaches 0, that is, the unit output is almost constant in history under this operating condition, the new write amount of the opportunity trace is directly set to zero to avoid the division by zero problem, and an upper limit for the write amount is set to limit the excessive impact of extreme values ​​on the trace space.

[0021] For example, assuming a generator unit has a base output of 92MW and an output bandwidth of 8MW, and its current actual output is 96MW, then the opportunity trace write amount is max(0,(92+8-96) / 8)=max(0,4 / 8)=0.5. This value indicates that this generator unit has 0.5 times the bandwidth of the upper limit of its ecological niche for additional generation space.

[0022] The risk trace is used to mark the system as being in a state of safety boundary or constraint tension. The triggering conditions are directly related to the unit's ramp exceeding the limit and the line power flow exceeding the limit. That is, if the ramp of a certain energy unit exceeds the preset ramp rate limit in the current period, or any transmission line associated with the unit exceeds the limit, the risk trace writing amount of the unit is directly set to 1.0; if none of the above conditions are met, the risk trace writing amount is directly set to 0.

[0023] The collaborative traces are used to reflect the collaborative behavior patterns of multiple units increasing their output, capturing the coordination and cooperation patterns of the units during the system load increase process. The writing rule is to obtain the normalized output increment of each unit. And statistics for the current time period. If the number of additional generating units reaches the preset minimum number of collaborative units, then for each additional generating unit, the collaborative trace write amount is defined as the product of its own normalized output increment and the average normalized output increment of other additional generating units; if the number of additional generating units does not reach the threshold, then the collaborative trace write amount of all units is set to zero. The normalized output increment It can be represented as: ; in, For the first The rated operating output of each energy unit.

[0024] For example, if the normalized output increment of Unit 1 is 0.01 and the normalized output increment of Unit 2 is 0.015 during a certain period, and there is no increase in wind power and photovoltaic output, then there are a total of 2 additional generating units, reaching the preset minimum threshold of 2 units. Therefore, the collaborative trace writing amount of Units 1 and 2 is 0.01 × 0.015 = 0.00015.

[0025] In one possible embodiment, the trace concentration matrix is ​​initialized as a 4×3 all-zero matrix. Assuming that in the first time period, only the three types of trace writes of Unit 1 are non-zero, and are 0.5, 0, and 0.0001 respectively, while the writes of the other units are all 0, the trace writes of Unit 1 are superimposed on the trace concentration matrix. Except for Unit 1, whose first column element is 0.5 and third column element is 0.0001, all other elements are 0. This matrix is ​​a slice of the trace tensor in the first time period. Before traversing the second time period, all elements of the trace concentration matrix are multiplied by a preset exponential decay factor, and then the trace writes of the second time period are obtained and superimposed to update all element values ​​of the trace concentration matrix. This process is repeated until the full-time simulation is completed. After traversing all T historical time periods, a complete historical trace tensor with a dimension of T×4×3 is obtained. At the same time, the trace concentration matrix after the end of the last time period is kept separately as the initial state of the trace space in the online stage, i.e., the initial trace matrix.

[0026] Step S14: Extract stable landform features from the trace tensor and simultaneously construct an initial rule template library. Specifically: First, a CP decomposition is performed on the trace tensor. Simultaneously, combined with preset stable terrain feature extraction rules, the trace tensor is approximated as the sum of R rank-one components. Each rank-one component is composed of the outer product of the time factor vector, energy factor vector, and trace factor vector, and each rank-one component corresponds to a typical scheduling behavior pattern, i.e., a stable terrain feature. The CP decomposition can be expressed as: ; For trace tensors; The decomposition rank, i.e., the number of extracted scheduling features, is determined using the cumulative explained variance criterion, i.e., from... We begin by gradually increasing the rank, calculating the explained variance of the decomposition results to the original tensor for each rank (i.e., the goodness of fit), and selecting the minimum value that makes the cumulative explained variance reach a preset threshold for the first time. Value as The final value; A time factor vector, used to describe the first... The activity intensity of each feature at different time periods, and its length is the number of all historical time periods traversed in step S13; Let be the energy factor vector, used to describe the _ The weight distribution of each feature on each energy unit, the length of which is the number of energy units participating in the scheduling; This is a trace factor vector used to describe the th trace factor vector. The relative proportions of the three types of traces mentioned above in each feature.

[0027] If the original data contains some missing values, such as invalid data for a certain unit during a certain period, the weighted least squares algorithm is used during decomposition to ignore the impact of missing values, so as to ensure the robustness of the decomposition results.

[0028] For example, after performing CP decomposition on the trace tensor, four rank-one components are obtained, which together constitute a stable geomorphic feature set. Each set of features corresponds to a typical scheduling behavior pattern. Assuming that the time factor of the first type of feature is active during the morning peak load ramp-up period, the energy factor has a high weight on the two thermal power units, and the proportion of collaborative traces in the trace factors is high, then it corresponds to the scheduling pattern of multi-unit collaborative power generation during the morning peak. The time factor of the second type of feature is active during windy weather, the energy factor has a negative weight on the thermal power units, and the proportion of opportunity traces in the trace factors is high, then it corresponds to the pattern of unit power generation reduction and margin improvement during the peak generation of new energy sources. The third type of feature corresponds to the risk scheduling pattern during line heavy load, and the fourth type corresponds to the smoothing pattern under midday photovoltaic fluctuations.

[0029] Next, for each set of stable landform features, the time periods in which the corresponding time factor vector values ​​are in the preset high percentile are selected, the mean vector of the multidimensional external condition vectors corresponding to these time periods is collected, and the mean vector is used as the situational pattern vector of the corresponding stable landform feature to represent the most typical applicable working condition of the scheduling rule. The energy factor vector is multiplied element by element with the comprehensive trace direction value to obtain the original direction vector. The original direction vector is then normalized to a preset symmetric interval, such as [-1,1], to generate the output correction ratio vector. In this vector, a positive value indicates that the corresponding unit is recommended to increase power generation, and a negative value indicates that it is recommended to reduce power generation. The absolute value indicates the relative intensity of the adjustment. The comprehensive trace direction value is generated based on the trace factor vector corresponding to the preset simulated trace, and simultaneously combined with preset rules. For example, it follows the physical logic that opportunity and synergy traces correspond to positive issuance increase and risk traces correspond to negative issuance decrease, and the comprehensive trace direction value is generated accordingly. Set as ,in , , These correspond to the values ​​of the opportunity, synergy, and risk trace factor vectors, respectively. The context pattern vector, the output correction ratio vector, and the preset initial confidence level are combined into a rule template. All rule templates together constitute the initial rule template library. At the same time, a KD tree spatial index structure is constructed based on the context pattern vector corresponding to each rule template, thereby realizing fast nearest neighbor query in multi-dimensional space.

[0030] Step S2: During the online operation phase, real-time operation data is collected according to a preset control cycle, a real-time external condition vector is constructed, and after a preset confidence gating judgment, the actual niche offset and niche construction potential of each energy unit are obtained, and the online trace concentration matrix and niche memory matrix are updated synchronously.

[0031] Specifically, step S2 includes steps S21-S23: Step S21: During the online operation phase, real-time operation data is collected according to a preset control cycle to construct a real-time external condition vector. Specifically: Obtain the latest real-time operating data, the data type of which is consistent with the previous system operating data, and construct the real-time external condition vector using the same process as step S11.

[0032] It should be noted that a rolling historical database is maintained during the online operation phase. This database always stores the most recent full-length running data with the same duration as the offline training data, and the following two update cycles are set: The first type is daily updates, which means that every day at dawn, the rolling historical database is used to re-execute the process of screening and statistically analyzing the nearest neighbor samples of the basic niche, refresh the internal nearest neighbor sample set of the base value function and the bandwidth function, and realize the adaptive update of the basic niche. The second type is quarterly updates, which means that every quarter, the mean and standard deviation of each dimension of the external condition vector are recalculated using the rolling database. Simultaneously, the restandardization of the rule templates and index updates are completed. At the same time, the context pattern vectors of all rule templates in the rule template library are reversed and restandardized using the updated new parameters, converted into standardized vectors under the new parameter system, and the KD tree index is reconstructed based on the updated context pattern vectors. Furthermore, to address the reference system shift issue caused by the basic niche update, after each daily update of the basic niche function, the latest base value function and bandwidth function are used to recalculate the actual niche shift for all time periods in the rolling historical database and replace the original previous shift sequence. This ensures that the shift used for historical significance determination during the online phase is in the same statistical caliber as the current niche function, thereby eliminating the judgment error caused by inconsistency in the reference system.

[0033] Step S22: After pre-set confidence gating, obtain the actual niche shift and niche construction potential of each energy unit. Specifically: First, compare the current real-time external condition vector of new energy penetration rate with the preset high proportion threshold value, and at the same time check whether the prediction confidence of wind power and photovoltaic power are both lower than the preset confidence lower limit. If both of the above conditions are met at the same time, the current period is marked as a low confidence period. During the low confidence period, due to insufficient data credibility, the system skips the offset calculation and trace writing operation. The actual niche offset directly uses the value of the previous period, and all new trace writing amounts are set to zero. Otherwise, the current time period is determined to be a normal time period.

[0034] Next, during normal periods, based on the current real-time external condition vector and the basic niche function, the output baseline and output bandwidth of each energy unit are obtained, and the actual niche offset for the current period is obtained. It can be represented as: .

[0035] It should be noted that, in order to solve the problem of offset jumps at the moment of niche update, this method performs an offset backtracking recalculation operation in the first control period after the daily niche update. That is, the offset of the previous period is recalculated using the updated new niche function, and the offset change of the current period is calculated based on the recalculated value, instead of directly using the old offset before the update. This eliminates the offset difference caused by the function update itself, so as to avoid false jumps in the potential.

[0036] Subsequently, based on the time-varying rate of change and historical significance components of the actual niche shift, corresponding niche building potentials are generated for each normal time period. These niche building potentials are normalized scalar values ​​used to measure the initiative and degree of anomalousness of the current energy unit's deviation from its basic niche; a higher value indicates a more significant deviation. Specifically: The time period change rate is used to measure the rate and direction of change of the actual ecological niche shift. It can be calculated by first calculating the difference between the current shift and the shift of the previous time period, then dividing by the current output bandwidth value. A very small constant can be added during the calculation to avoid division by zero. Finally, the result is mapped to [0,1] by the hyperbolic tangent function. The historical significance component is used to determine whether the current actual niche shift is at an abnormal level in similar historical conditions. This can be achieved by filtering all historical samples from the previous shift sequence whose Euclidean distance between the previous external condition vector and the current real-time external condition vector is less than a preset threshold, and calculating the mean and standard deviation of the actual niche shift of these samples. If the absolute value of the difference between the current actual niche shift and the historical mean exceeds the preset threshold, a significant shift anomaly is determined, and the value of the historical significance component is directly set to 1.0; otherwise, it is directly set to 0. The two indicators mentioned above are weighted and summed according to the weighting coefficients set by domain experts in advance to generate the niche building potential.

[0037] For example, if a unit's current actual niche offset is 6MW, the previous period's actual niche offset was 2MW, and the current output bandwidth is 8MW, then the difference in actual niche offset is 4MW. Dividing this by the output bandwidth gives 0.5, which, after mapping using the tanh function, is approximately 0.462, meaning the period's change rate is 0.462. Assuming the mean of actual niche offset under similar historical operating conditions is 1MW and the standard deviation is 2MW, the difference between the current actual niche offset and the historical mean is 5MW, exceeding the preset threshold. Therefore, the historical significance component is 1.0. After weighted summation according to preset weight coefficients, the niche building potential is 0.6×0.462+0.4×1.0≈0.68, which is at a moderately high level. This indicates that the unit's current output deviates from the historical normal state and changes rapidly.

[0038] Step S23, update the online trace concentration matrix and niche memory matrix, specifically: First, the system maintains an online trace concentration matrix in memory. Its initial state is assigned by the initial trace matrix generated in the offline stage. At the end of each control cycle, all elements of the online trace concentration matrix are multiplied by a decay factor that is completely consistent with the offline simulation. For normal periods, according to the same rules used in step S1, the writing volume of the three types of traces in the current period is obtained, the current writing matrix is ​​generated, and the elements are superimposed onto the online trace concentration matrix. For low-confidence periods, the current write matrix is ​​directly set to an all-zero matrix. After being superimposed on the online trace concentration matrix, the trace concentration in the matrix does not change.

[0039] Next, for each pair of energy units A sliding time window is constructed by taking the previous historical time period, including the current time period. The actual output sequence of the two energy units within this sliding time window is extracted, and the Pearson correlation coefficient is calculated. Then, the elements of the niche memory matrix are updated using the exponential moving average formula, i.e. ; in A smoothing coefficient is pre-set for domain experts to ensure the smoothness and continuity of the matrix.

[0040] It should be noted that the niche memory matrix is ​​an N×N symmetric matrix, where N is the number of energy units participating in the scheduling, and its elements... Used to characterize Time-of-use energy unit and The initial value of this matrix is ​​determined by calculating the global Pearson correlation coefficient between any two energy units using offline, full-time historical output data.

[0041] Step S3: Generate online trace vectors based on the online trace concentration matrix, and match them with the basis vectors corresponding to each group of stable landform features, outputting the best matching landform feature or no matching flag.

[0042] Specifically, step S3 includes steps S31-S32: Step S31: Generate online trace vectors based on the online trace concentration matrix, and obtain the basis vectors corresponding to each group of stable landform features. Specifically: First, the online trace concentration matrix is ​​flattened into a one-dimensional online trace vector according to a preset order. For example, the online trace concentration matrix is ​​expanded in the order of "unit number first, then trace type". Assuming that the online trace concentration matrix is ​​4 rows and 3 columns for a certain period, the trace concentrations of the three types of units are 0.785, 0, and 0.0002 for unit 1, 0.62, 0, and 0.00018 for unit 2, 0.3, 0, and 0 for wind turbine unit 3, and 0.45, 0, and 0 for energy storage unit 4. Then the flattened online trace vector can be represented as follows: .

[0043] Next, the energy factor vector and trace factor vector corresponding to the stable landform features are subjected to the Kronecker product operation, and then flattened in the same unfolding order as the online trace concentration matrix to generate the corresponding basis vector.

[0044] For example, given an energy factor vector b_r=[0.8,1.0,-0.3,-0.5] and a trace factor vector c_r=[0.7,0.1,0.8] for a certain landform feature, after the Kronecker product operation, the first three elements of the basis vector are 0.8×0.7=0.56, 0.8×0.1=0.08, and 0.8×0.8=0.64, corresponding to the three types of trace weights for Unit 1; the next three elements are 1.0×0.7=0.7, 1.0×0.1=0.1, and 1.0×0.8=0.8, corresponding to the three types of trace weights for Unit 2, and so on, finally yielding a 12-dimensional basis vector. This vector fully describes the ideal relative proportions of various types of traces for each unit under this landform pattern.

[0045] Step S32: Match the online trace vector with the basis vectors corresponding to each group of stable landform features, and output the best matching landform feature or no matching flag. Specifically: For each set of stable landform features, calculate the morphological similarity between the current online trace vector and the corresponding basis vector. The morphological similarity is measured using cosine similarity with absolute value, i.e.: ; in, This is the current online trace vector; For the first Basis vectors of stable landform features; This is a preset minimum positive number, used to prevent abnormal values ​​where the denominator is zero.

[0046] Traverse all stable landform features and obtain the morphological similarity corresponding to each stable landform feature. Select morphological similarity The most stable landform feature is used as the candidate matching result, and its feature number and morphological similarity are recorded simultaneously. The value; The highest morphological similarity If the value is not lower than the preset matching threshold, the matching is considered successful. The corresponding stable landform feature number, the corresponding stable landform feature and morphological similarity are output. The output stable landform feature is defined as the optimal matching landform feature, and the process proceeds to step S4. If the highest morphological similarity If the value is lower than the preset matching threshold, it is determined that the current trace spatial morphology cannot be effectively explained by any known landform features, outputs no matching flag, and directly jumps to the conservative scheduling branch of step S5 without performing active output correction.

[0047] Step S4: When a match is successful, the corresponding rule template is retrieved from the initial rule template library based on the optimal matching geomorphic features. Simultaneously, the current working condition base value, matching similarity and niche construction potential are combined to obtain the absolute output correction amount of each energy unit. After being limited by the basic niche boundary, the final value of the correction amount is output.

[0048] Specifically, step S4 includes steps S41-S43: Step S41: Retrieve the corresponding rule template from the initial rule template library based on the optimal matching terrain features. Specifically: This method employs a two-level retrieval mechanism to achieve accurate matching of each template in the initial rule template library. Each template in the initial rule template library is associated with a corresponding stable landform feature through a feature number. The two-level retrieval mechanism is as follows: The first level is feature index filtering: using the number of the best matching landform feature output in step S3 as the first-level index, all templates associated with the landform feature are filtered out from the initial rule template library, thereby selecting candidate templates.

[0049] The second level is the nearest neighbor matching of working conditions: for the selected candidate templates, the Euclidean distance between the current real-time external condition vector and the corresponding situation mode vector of each candidate template is obtained, and the template with the smallest distance is selected as the final rule template to be called.

[0050] Step S42: Combining the current operating condition baseline, matching similarity, and niche construction potential, obtain the absolute output correction for each energy unit. Specifically: The corresponding output correction ratio vector is read from the final rule template. Combined with the morphological similarity output in step S3, the current operating condition baseline value, and the niche construction potential, the output correction ratio vector is converted into a relative correction ratio to generate an absolute output correction amount. It can be represented as: ; in, The highest morphological similarity output in step S3 The value; This is the output correction ratio vector for energy unit i in the final rule template; Energy unit under current operating conditions The base output value is the preset value; The pre-set potential gain coefficient for domain experts is used to fine-tune the niche construction potential.

[0051] Step S43: After applying basic niche boundary limiting to the absolute output correction, output the final value of the correction. Specifically: The absolute output correction is applied to a limiting stage using a truncation function and a basic niche output range, and the final value of the correction is output, i.e.: If the absolute output correction is within the basic niche output range, no action is taken. If the absolute output correction is lower than the lower bound of the basic niche output range, its value is directly limited to the lower bound of the basic niche output range. Similarly, if the absolute output correction exceeds the upper limit of the basic niche output range, its value is directly limited to the upper limit of the basic niche output range.

[0052] Step S5: Perform safety constraint verification on the tentative output command after superimposed correction final value. After the verification is passed, issue a formal scheduling command, collect the actual operation data after the command is executed, and dynamically update the confidence level of the corresponding rule template in combination with the comprehensive feedback factor.

[0053] Specifically, step S5 includes steps S51-S53: Step S51: Generate a trial output command and simultaneously perform a safety constraint check. Specifically: First, the final value of the correction for each energy unit is superimposed on the actual output for the current period to generate the corresponding trial output command.

[0054] Next, the following safety constraint checks are performed on the tentative output command: Power balance verification is to verify whether the deviation between the total power output of the entire system and the total load of the system is within the preset safety range. Unit ramp-up constraint verification verifies whether the output change of each energy unit is within the corresponding preset ramp-up rate limit range, which is set by domain experts based on the physical characteristics of the unit. Line power flow safety verification involves verifying whether the power flow of all transmission lines is within the preset safety capacity range. For example, based on the DC power flow model, the power flow value of each line is calculated in conjunction with the trial output command, and it is verified whether the absolute value of the power flow of all lines does not exceed the line rated capacity multiplied by the preset safety margin coefficient. This safety margin coefficient is set by experts in the field based on actual conditions.

[0055] Next, if any of the above safety constraints fails the check, an iterative correction process is initiated. The correction amount is adjusted specifically according to the type of constraint violation until all constraints are satisfied or the maximum number of iterations is reached. Specifically: If the power balance check fails, perform iterative correction as follows: Step A1: Obtain the total system imbalance and determine the direction of the deviation. Specifically: Calculate the difference between the sum of the trial outputs of all units and the total system load, i.e., the total imbalance. ; in Energy unit in the exploratory power output command The actual output value; like If so, it is determined that the power generation output is excessive and a power reduction adjustment needs to be implemented; like If so, it is determined to be a power generation deficit, and additional power generation adjustments are required.

[0056] Step A2: Filter and sort the currently adjustable generating units. Specifically: For the direction of deviation, obtain the remaining adjustable capacity corresponding to each unit; In scenarios where additional share issuance is required, the remaining upward adjustment capacity of a single unit shall be the smaller of the two: "the remaining upward adjustment space allowed by the ramp rate" and "the difference between the upper limit of the basic niche and the current trial output". In scenarios where power generation reduction adjustments are required, the remaining capacity reduction for a single unit is the smaller of "the remaining reduction space allowed by the ramp rate" and "the difference between the current trial output and the lower limit of the basic ecological niche". Select units with remaining adjustable capacity greater than 0, and sort them in descending order of the absolute value of the current correction amount.

[0057] Step A3: Based on the remaining adjustable capacity of each unit, offset the total system imbalance unit by unit. Specifically: According to the sorting results obtained in step A2, the remaining adjustable capacity of the corresponding unit is taken in sequence, the output value of the unit in the trial output command is adjusted in the same direction, and the corresponding value of the remaining imbalance is deducted simultaneously. After each unit is adjusted, immediately verify whether the remaining total system imbalance is within the allowable deviation range of power balance. If, after adjustment to a certain unit, the absolute value of the remaining total system imbalance is less than or equal to the product of the balance tolerance and the total system load, then the power balance correction is complete, and the adjusted trial output command is retained. If, after traversing all adjustable units, the remaining total system imbalance still exceeds the allowable range, the power balance adjustment is deemed to have failed, and the system proceeds to the fallback logic for the maximum number of iterations.

[0058] If the unit ramping constraint verification or line power flow safety verification fails, all relevant units involved in the over-limit are traversed and sorted in descending order by the absolute value of the corresponding force of each unit. Each time, the correction amount of the unit with the largest over-limit is reduced by half, the trial command is regenerated and the full set of safety constraint verification is executed again.

[0059] It should be noted that if a correction scheme that satisfies all constraints is found within the preset maximum number of iterations, it will be directly adopted and the adjusted instructions will be issued. If it still cannot pass all the above safety constraint checks after the preset maximum number of iterations, all corrections in this round will be abandoned and a conservative strategy will be adopted, that is, the scheduling instructions of the previous period will be directly issued to ensure the safe operation of the system.

[0060] Step S52: Collect actual operating data after command execution and generate a comprehensive feedback factor. Specifically: After the formal scheduling command is issued and executed, the actual operating data corresponding to the next control cycle is collected, and the following preset performance indicators are extracted: Response accuracy is used to measure the degree of agreement between the actual output change of a unit and the expected correction. It is obtained by dividing the absolute value of the difference between the actual output change and the expected correction of each unit by the absolute value of the expected correction, and the response accuracy is defined as 1 minus the average value of the response deviation ratios of all units.

[0061] The constraint safety level is used to characterize the severity of line overruns during a given time period. It can be defined as the ratio of 1 minus the number of overrunning lines to the total number of lines.

[0062] The above performance indicators are weighted and fused to generate a corresponding comprehensive feedback factor. The weighting coefficients of the weighting and fusion are set by domain experts according to actual needs.

[0063] Step S53: Dynamically update the confidence level of the corresponding rule template based on the comprehensive feedback factor. Specifically: If the comprehensive feedback factor is greater than or equal to the preset high feedback threshold, the rule is judged to be of excellent performance, and its confidence level is increased by a preset reward value, such as 5 percentage points, but the maximum is set to not exceed 100%. If the comprehensive feedback factor is less than or equal to the preset low feedback threshold, the judgment rule is considered to be performing poorly, and its confidence level is reduced by the preset penalty magnitude, such as by 5 percentage points, but a minimum confidence level of no less than 10% is set.

[0064] If the confidence level of a rule template decays below the lower limit, the rule is deemed to have no practical value and is removed from the rule template library.

[0065] It should be noted that the above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0066] It should be understood that the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent three cases: A alone, A and B simultaneously, and B alone. A and B can be singular or plural. Additionally, the character " / " in this article generally indicates an "or" relationship between the preceding and following related objects, but it can also represent an "and / or" relationship. Please refer to the context for a more accurate understanding.

[0067] It should be understood that, in the embodiments of the present invention, the order of the above-mentioned process numbers does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0068] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for extracting dispatch rules for a multi-energy complementary intelligent power system, characterized in that, It includes the following steps: S1: Based on past system operation data, construct past external condition vector sequences, establish basic ecological niche functions and generate past offset sequences, define simulated trace accumulation process and generate trace tensors and initial trace matrices, extract stable geomorphic features from trace tensors, and simultaneously construct an initial rule template library. S2: During the online operation phase, real-time operation data is collected according to the preset control cycle, and a real-time external condition vector is constructed. After the preset confidence gating judgment, the actual niche offset and niche construction potential of each energy unit are obtained, and the online trace concentration matrix and niche memory matrix are updated synchronously. S3: Generate online trace vectors based on the online trace concentration matrix, and match them with the basis vectors corresponding to each group of stable landform features, outputting the best matching landform feature or no matching flag; S4: When a match is successful, the corresponding rule template is retrieved from the initial rule template library based on the optimal matching geomorphic features. Simultaneously, the current working condition base value, matching similarity and niche construction potential are combined to obtain the absolute output correction amount of each energy unit. After being limited by the basic niche boundary, the final value of the correction amount is output. S5: Perform safety constraint verification on the tentative output command after superimposed correction final value. After the verification is passed, issue formal scheduling command, collect actual operation data after command execution, and dynamically update the confidence level of the corresponding rule template in combination with comprehensive feedback factor.

2. The method for extracting scheduling rules for a multi-energy complementary intelligent power system according to claim 1, characterized in that, Step S1 includes: For each historical period in the past system operation data, a multidimensional external condition vector is constructed. The multidimensional external condition vector includes at least the system load rate, the new energy penetration rate, the time period label, and the seasonal coefficient. The system load factor is expressed as the ratio of the total system load to the rated total system load. The renewable energy penetration rate is expressed as the ratio of the sum of the predicted wind power and the predicted photovoltaic power to the total system load. Time period labels are represented as the time period scalar values ​​corresponding to the time periods after mapping each time point within a day to a preset continuous numerical range; The seasonal coefficient is represented by dividing the year into several segments by season or month and assigning different discrete or continuous values ​​to reflect seasonal differences. Standardize the multidimensional external condition vectors of all past periods dimension by dimension to generate a sequence of past external condition vectors.

3. The method for extracting scheduling rules for a multi-energy complementary intelligent power system according to claim 2, characterized in that, The method further includes: The basic niche function includes a base function and a bandwidth function; For each energy unit, retrieve the K time periods with the smallest Euclidean distance to the standardized multidimensional external condition vector from the previous external condition vector sequence; Extract the actual output value of the energy unit within the first K time periods, and use the corresponding median as the output of the base function and the corresponding standard deviation as the output of the bandwidth function; Based on the base function and bandwidth function, the basic niche output range of the energy unit under the corresponding operating condition is obtained. The basic niche output range is represented as [base function output value - bandwidth function output value, base function output value + bandwidth function output value]; For each energy unit in each past period, the corresponding actual niche offset is obtained. The actual niche offset is represented as the difference between the actual output value of the energy unit in the past period and the output value of the base function under the corresponding operating condition. The actual niche shift of each energy unit in all past periods constitutes the past shift sequence.

4. The method for extracting scheduling rules for a multi-energy complementary intelligent power system according to claim 2, characterized in that, The method further includes: Initialize a trace concentration matrix of a preset size, iterate through all past periods, and in each past period, perform evaporation decay on the trace concentration matrix based on a preset exponential decay factor; Obtain the amount of writing of the preset simulated traces within the corresponding time period and add them to the trace concentration matrix element by element, and save the updated trace concentration matrix as a complete slice; The preset simulation traces include at least opportunity traces, risk traces, and collaboration traces; After traversing all previous periods, the corresponding trace tensor and initial trace matrix are generated; Perform CP decomposition on the trace tensor, and simultaneously combine it with preset stable landform feature extraction rules to approximate the trace tensor as the sum of multiple rank-one components; The rank-one component is the stable landform feature, and it is composed of the outer product of the time factor vector, the energy factor vector, and the trace factor vector. An initial rule template library is constructed based on the stable landform features.

5. The method for extracting scheduling rules for a multi-energy complementary intelligent power system according to claim 4, characterized in that, An initial rule template library is constructed based on the aforementioned stable landform features, including: For each set of stable landform features, the time period in which the corresponding time factor vector value is in the preset high percentile is selected, and the mean vector of the multidimensional external condition vector corresponding to the selected time period is used as the context pattern vector of the corresponding stable landform feature. Based on the energy factor vector and the comprehensive trace direction value, the original direction vector is obtained, and the original direction vector is normalized to a preset symmetric interval to generate the output correction ratio vector. The comprehensive trace direction value is generated based on the trace factor vector corresponding to the preset simulated trace, and simultaneously combined with preset rules. The scenario pattern vector, the output correction ratio vector, and the preset initial confidence level are combined into a rule template, and all rule templates together constitute the initial rule template library.

6. The method for extracting scheduling rules for a multi-energy complementary intelligent power system according to claim 1, characterized in that, Step S2 includes: Real-time operational data is collected in each preset control cycle to construct a real-time external condition vector; Based on the real-time external condition vector, and combined with the preset confidence gating, the confidence of the time period is determined. The time period that meets the preset low confidence condition is determined as a low confidence time period, and otherwise it is determined as a normal time period. For low-confidence periods, the actual niche offset value corresponding to the previous period is directly used as the actual niche offset value for the current period, and the online trace concentration matrix is ​​not updated. For normal periods, the corresponding actual niche offset is generated based on the actual output value of the energy unit in the current period and the basic niche function under the current operating conditions. Based on the time-period variation rate and historical significance component of the actual niche shift, a corresponding niche construction potential is generated for each normal time period. The current online trace concentration matrix and niche memory matrix are updated according to the preset trace concentration update rules and preset niche update functions.

7. The method for extracting scheduling rules for a multi-energy complementary intelligent power system according to claim 1, characterized in that, Step S3 includes: The online trace concentration matrix is ​​flattened in a preset order to generate an online trace vector; For each set of stable geomorphic features, a basis vector is generated by combining the energy factor vector and trace factor vector contained within the stable geomorphic features. Based on a preset similarity metric, similarity matching is performed between the online trace vector and the basis vectors of each stable landform feature; If the highest similarity between the online trace vector and the basis vector of each stable landform feature is not lower than the preset matching threshold, then the stable landform feature with the highest similarity is directly used as the optimal matching landform feature. Otherwise, output a no-match flag and execute the preset conservative scheduling strategy.

8. The method for extracting scheduling rules for a multi-energy complementary intelligent power system according to claim 1, characterized in that, Step S4 includes: When a match is successful, the corresponding rule template is selected from the initial rule template library using the best matching geomorphic feature as an index to generate a candidate rule template set; The candidate rule template set is filtered based on preset filtering indicators to obtain the optimal rule template corresponding to the current working condition. By combining matching similarity, current operating condition baseline, and niche building potential, the absolute output correction of each energy unit is obtained; The absolute output correction is superimposed on the current actual output, and after being processed by the basic niche boundary limit, the final value of the correction is generated.

9. The method for extracting scheduling rules for a multi-energy complementary intelligent power system according to claim 1, characterized in that, Step S5 includes: A safety constraint check is performed on the tentative output command after the final value of the superimposed correction. The safety constraint check includes at least power balance check, unit ramping constraint check and line power flow safety check. If any check fails, iterative correction is performed in accordance with preset rules until all security constraint checks are passed or the preset maximum number of iterations is reached. If all safety constraints still cannot be passed after reaching the preset maximum number of iterations, then abandon this round of correction and issue the formal scheduling instruction of the previous period. After the formal scheduling instruction is issued and executed, the actual operation data of the next cycle is collected and combined with the preset performance indicators to generate a comprehensive feedback factor; The confidence level of the rule template used in this call is dynamically updated based on the comprehensive feedback factor.

Images