A method and system for determining key state characteristics of a power grid

Abstract

The present invention discloses a method and system for determining key state characteristics of a power grid, the method comprising: calculating state characteristic values ​​in a preset feature set based on acquired power grid operation data; training using a multi-task time series learning model based on all state characteristic values ​​in the feature set to obtain predicted values ​​of power grid performance and key state characteristics corresponding to the predicted values ​​of power grid performance; the multi-task time series learning model not only realizes the prediction of the future operation state of the power grid, but also can determine the key state characteristics corresponding to the predicted values ​​of power grid performance, which helps control personnel to accurately grasp the current power grid state and development trend through the key state characteristics, improves the intelligence of power grid control business, and verifies the accuracy of the future operation state of the power grid by observing the changes in the key state characteristics of the power grid, which is conducive to timely detection of faults.

Description

Technical Field

[0001] The present invention relates to the field of power system operation and control, and in particular to a method and system for determining key state characteristics of a power grid. Background Art

[0002] With the rapid development and engineering practice of ultra-high voltage AC and DC large-capacity transmission and new energy, the scale of the power grid has become extremely large and the operation mode of the power grid has become extremely complex. The intermittent and random nature of new energy and distributed power sources has also brought about problems that are difficult to accurately predict. These have led to a significant increase in the complexity of regulation and control business, and have put forward new requirements for power grid dispatching, control and operation.

[0003] Ideal grid dispatching uses historical operation data to analyze and evaluate the ideal state of grid operation, find out the gap between the historical operation state and the ideal state of the grid, conduct "re-dispatch" and "reflection", and thus improve the level of grid dispatching and operation control. The realization of ideal grid dispatching is based on the accurate analysis and evaluation of a large amount of historical data. Traditional analysis and evaluation methods based on deterministic models are difficult to meet its needs. At the same time, when judging the current state and development trend of the grid, the dispatchers need to consider multiple state characteristics, which leads to high workload and low efficiency, and also requires high skills of the dispatchers. Summary of the invention

[0004] In order to solve the above-mentioned deficiencies in the prior art, the present invention provides a method and system for determining the key state characteristics of a power grid. The determination method provided by the present invention utilizes massive historical power grid data, and on the basis of traditional dispatching decision-making methods, predicts the future operating state of the power grid and explores the key factors affecting the operation of the power grid. The method first sorts out the operating characteristics of the power grid from the three aspects of safety, economy, and energy saving, and combines the self-learning algorithm based on LSTM to build a multi-task sequence learning method model, which realizes the prediction of the future operating state of the power grid and the extraction of key state characteristics of the power grid, which helps to help control personnel accurately grasp the current power grid status and development trends, and improves the intelligence of power grid control services.

[0005] The present invention provides a method for determining key state characteristics of a power grid, comprising:

[0006] Calculating state characteristic values ​​in a preset characteristic set based on the acquired power grid operation data;

[0007] Based on all state feature values ​​in the feature set, a multi-task time series learning model is used for training to obtain a predicted value of power grid performance and key state features corresponding to the predicted value of power grid performance;

[0008] Among them, the multi-task time series learning model is constructed based on the relationship between state features and power grid performance; the state features in the feature set are extracted based on power grid security, economy and energy saving.

[0009] Preferably, the multi-task time series learning model is used for training based on all state feature values ​​in the feature set to obtain the predicted value of power grid performance and the key state features corresponding to the predicted value of power grid performance, including:

[0010] Obtain historical performance scores of power grid in chronological order;

[0011] The time series and the historical performance scores of the power grid corresponding to the time series and the state characteristic values ​​are sequentially input into the multi-level discrete wavelet transform module and the Seq2Seq predictor to obtain the predicted value of the power grid performance;

[0012] Input the time series and the historical performance scores of the power grid corresponding to the time series and the state characteristic values ​​into the correlation module and the Seq2Seq predictor in sequence to obtain the correlation coefficients of all state characteristics;

[0013] Sorting the correlation coefficients of all state characteristics, and selecting the key state characteristics of the power grid corresponding to the predicted values ​​of the power grid performance in order;

[0014] The Seq2Seq predictor includes multiple LSTM encoder-decoders and hidden layers.

[0015] Preferably, the time series and the power grid historical performance score and state characteristic value corresponding to the time series are sequentially input into a multi-level discrete wavelet transform module and a Seq2Seq predictor to obtain a predicted value of power grid performance, including:

[0016] The time series and the historical performance scores of the power grid corresponding to the time series and the state characteristic values ​​are input into a multi-level discrete wavelet transform module to decompose the input time series into multiple subsequences;

[0017] The decomposed subsequences are fed into the Seq2Seq predictor to output the predicted value of power grid performance.

[0018] Preferably, the time series and the historical performance score of the power grid corresponding to the time series and the state characteristic value are sequentially input into the correlation module and the Seq2Seq predictor to obtain the correlation coefficients of all state characteristics, including:

[0019] Input the time series and the historical performance scores of the power grid corresponding to the time series and the state characteristic values ​​into the correlation module;

[0020] Obtaining a time series of correlation coefficients of each state feature in different time periods based on the correlation module;

[0021] Feed the time series of all state features into the stacked LSTM module;

[0022] The output of the stacked LSTM module is input into the hidden layer of the Seq2Seq predictor, and the output of the hidden layer is input into a fully connected neural network to obtain the correlation coefficients of all state features.

[0023] Preferably, the power grid historical performance score is obtained by expert scoring or setting weights.

[0024] Preferably, the state features extracted based on security include:

[0025] N-1 pass rate, N-2 pass rate of main sections, short-circuit current, power flow operation, main transformer safety, line safety, grid operation balance, rotating reserve deficiency rate, supply area load reactive power margin, voltage safety, frequency qualification rate, frequency quality rate and load forecast accuracy.

[0026] Preferably, the state characteristics extracted based on economy include:

[0027] The average electricity purchase cost deviation rate, network loss rate, rotating reserve excess rate, average load rate of peak load units and reactive power balance degree of the supply area.

[0028] Preferably, the state features extracted based on energy saving include:

[0029] Average coal consumption for power generation and equivalent average load factor of generating units.

[0030] Based on the same inventive concept, the present invention also provides a system for determining key state characteristics of a power grid, comprising:

[0031] A calculation module, used for calculating the state characteristic value in the preset characteristic set based on the acquired power grid operation data;

[0032] A result module is used to train a multi-task time series learning model based on all state feature values ​​in the feature set to obtain a predicted value of power grid performance and key state features corresponding to the predicted value of power grid performance;

[0033] Among them, the multi-task time series learning model is constructed based on the relationship between state features and power grid performance; the state features in the feature set are extracted based on power grid security, economy and energy saving.

[0034] Preferably, the result module includes:

[0035] A scoring unit, used to obtain the historical performance score of the power grid in time series;

[0036] A prediction unit is used to input the time series and the historical performance score of the power grid corresponding to the time series and the state characteristic value into the multi-level discrete wavelet transform module and the Seq2Seq predictor in sequence to obtain the predicted value of the power grid performance;

[0037] A coefficient unit, used to input the time series and the historical performance score of the power grid corresponding to the time series and the state characteristic value into the correlation module and the Seq2Seq predictor in sequence to obtain the correlation coefficients of all state characteristics;

[0038] A result unit, used to sort the correlation coefficients of all state characteristics, and select the key state characteristics of the power grid corresponding to the predicted value of the power grid performance in order;

[0039] The Seq2Seq predictor includes multiple LSTM encoder-decoders and hidden layers.

[0040] Compared with the prior art, the present invention has the following beneficial effects:

[0041] The technical solution provided by the present invention calculates the state characteristic values ​​in a preset feature set based on the acquired power grid operation data; trains using a multi-task time series learning model based on all the state characteristic values ​​in the feature set to obtain the predicted value of the power grid performance and the key state characteristics corresponding to the predicted value of the power grid performance; the multi-task time series learning model not only realizes the prediction of the future operation state of the power grid, but also can determine the key state characteristics corresponding to the predicted value of the power grid performance, which helps the control personnel to accurately grasp the current power grid state and development trend through the key state characteristics, improves the intelligence of the power grid control business, and verifies the accuracy of the future operation state of the power grid by observing the changes in the key state characteristics of the power grid, which is conducive to timely detection of faults. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 A flow chart of a method for determining key state characteristics of a power grid provided by the present invention;

[0043] Figure 2 A schematic diagram of the structure of the multi-task time series learning model provided by the present invention. DETAILED DESCRIPTION

[0044] In order to better understand the present invention, the content of the present invention is further described below in conjunction with the accompanying drawings and examples.

[0045] like Figure 1 As shown, the present invention provides a method for determining key state characteristics of a power grid, comprising:

[0046] S1 calculates the state characteristic values ​​in the pre-set characteristic set based on the acquired power grid operation data;

[0047] S2: training using a multi-task time series learning model based on all state feature values ​​in the feature set to obtain a predicted value of power grid performance and key state features corresponding to the predicted value of power grid performance;

[0048] Among them, the multi-task time series learning model is constructed based on the relationship between state features and power grid performance; the state features in the feature set are extracted based on power grid security, economy and energy saving.

[0049] S1 calculates the state characteristic values ​​in the pre-set feature set based on the acquired power grid operation data, specifically refers to the extraction of the main state characteristics of the power grid, and forms feature sets of different themes from the aspects of power grid security, economy, energy saving, etc.;

[0050] The present invention needs to select an artificial intelligence algorithm suitable for extracting key state characteristics of the power grid and build a learning model based on the intelligent algorithm;

[0051] S2 uses a multi-task time series learning model to train based on all the state feature values ​​in the feature set to obtain the predicted value of the power grid performance and the key state characteristics corresponding to the predicted value of the power grid performance, specifically referring to the self-learning and self-refining of the key state characteristics of the power grid based on the artificial intelligence algorithm model.

[0052] This embodiment specifically analyzes the solution of the present invention:

[0053] (1) Extract the grid status feature set from the aspects of grid security, economy, energy saving, etc., including:

[0054] As the scale of power grid continues to expand, the complexity of power grid operation continues to increase, which has brought great challenges to my country's traditional power grid dispatching and operation management methods. It is urgent to scientifically and comprehensively grasp the power grid operation characteristics and comprehensively improve the dispatching management level. The power grid operation feature set is sorted out based on the actual needs of dispatchers and the dispatching needs of the actual power grid.

[0055] 1) N-1 pass rate (x) 1_1 )

[0056] The N-1 pass rate reflects whether the grid dispatching plan can meet the requirements of the first-level safety and stability standards. The N-1 pass rate should generally reach 100%. It is not ruled out that the pass rate may be lower than 100% in areas with particularly weak grid structures. Stability control measures need to be taken to meet the requirements.

[0057] 2) Main section N-2 pass rate (x 1_2 )

[0058] The N-2 pass rate of the main section is used to examine the strength of the power grid structure and the degree of realization of the second-level safety and stability standards. The "2" here refers to double-circuit lines on the same tower and parallel double-circuit lines. The N-2 pass rate is calculated based on the assumption that the power grid can remain stable without taking stability control measures. There is no specific requirement for its value in the regulations and guidelines.

[0059] 3) Short circuit current (x 1_3 )

[0060] As the density of power grids increases, the problem of excessive short-circuit current becomes increasingly prominent and becomes a key concern in power system operation.

[0061] In order to assess whether there is a problem of exceeding the short-circuit current standard in some operation modes in the day-ahead plan, the short-circuit current index is formulated:

[0062]

[0063]

[0064] Where: I i,k I is the larger value of the unidirectional short-circuit current and the three-phase short-circuit current on bus i; i,k,max is the maximum breaking current of the switchgear; P k is the short-circuit current level under the kth operation mode; Ω is the set of operation modes determined by the day-ahead plan; n e is the number of buses; n ey is the number of buses with excessive short-circuit current; a ei is the weight factor of busbar i.

[0065] 4) Flow operation (x 1_4 )

[0066] During the dispatching process, the dispatching department needs to control the flow of the main sections, timely grasp the flow conditions of the sections, and avoid the situation where the flow of the sections is overloaded or exceeds the limit.

[0067]

[0068] Where: P si is the actual flow of section i at the current statistical point; P si.max is the transmission limit of section i; n s is the number of main sections; n sy is the number of sections where the tidal current exceeds the limit; α si It is the artificially determined section importance weight.

[0069] 5) Main transformer safety (x 1_5 )

[0070] The main transformer load rate is used to judge the safety of equipment operation. If the load rate is too high (heavy load or overload), it means that there are certain hidden dangers in the safe operation of the equipment.

[0071]

[0072] Where: S i is the apparent power of transformer i at the statistical moment; S i.max is the capacity of transformer i; n t is the number of transformers; n ty is the number of transformers that are out of limit; α ti is the artificially determined importance weight of transformer i.

[0073] 6) Line safety (x 1_6 )

[0074] The line load rate is used to judge the safety of line operation. If the load rate is too high (heavy load or overload), it means that there are certain hidden dangers in the safe operation of the line.

[0075]

[0076] Where: I i is the current value of line i at the statistical moment; I i.max is the current limit of line i; n l is the number of lines; n ly is the number of lines with over-limit conditions; α li is the artificially determined importance weight of line i.

[0077] 7) Grid operation balance (x 1_7 )

[0078] The grid operation balance reflects the comprehensive level of grid power flow operation and has global characteristics. Under the same load level, the smaller the value, the more balanced the power flow of each line of the grid, and the higher the safety of grid operation.

[0079]

[0080] R li =I li / I li.max (7)

[0081] Where: R li is the load rate of line i; R l.ave is the average load rate of the line; I li is the current value of line i at the current statistical point; I li.max is the current limit of line i; n l The number of lines.

[0082] 8) Spinning reserve deficiency rate (x) 1_8 )

[0083] According to the basic regulations in this field, the spinning reserve ratio should be controlled at 2% to 5%. In order to reflect the phenomenon of insufficient spinning reserve during the previous day's operation, the spinning reserve shortage ratio index is now used as follows:

[0084] Spinning reserve deficiency rate = max(positive reserve deficiency rate, negative reserve deficiency rate) (8)

[0085]

[0086] 9) Reactive power margin of supply area load (x 1_9 )

[0087] According to the requirements of reactive power zoning balance, the nodes in the supply area are strongly coupled, and the electrical coupling between supply areas is weak. First, the reactive power supply in the supply area should meet the reactive power demand of the load in the supply area in the current state, and there should be sufficient reactive power reserves to meet the reactive power demand of the load in the supply area within a certain range. Therefore, the reactive power margin of the supply area load is defined as follows:

[0088]

[0089] Where: Q Li is the total reactive load in zone i; Q Gi The reactive power compensation capacity provided for zone i.

[0090] 10) Voltage safety (x 1_10 )

[0091] Maintaining voltage stability is a basic requirement for power system operation and dispatch. Evaluate voltage conditions;

[0092]

[0093] Where: V i is the voltage value of bus i at the statistical moment; V i,max 、V i,min are the upper and lower limits of bus voltage respectively; R Vi is the voltage deviation of bus i at the statistical moment; n e is the total number of busbars; n ey is the number of busbars with voltage exceeding the limit; α Vi is the artificially determined importance weight of bus i.

[0094] 11) Frequency pass rate (x) 1_11 )

[0095] The basic requirement for power grid operation is to ensure active power balance, that is, the active power generated by all power plants is equal to the sum of the active power consumed by all users and network losses and plant power consumption.

[0096]

[0097] Where: T is the total running time; T 0.2 This is the time when the frequency deviation exceeds ±0.2Hz.

[0098] 12) Frequency quality rate (x 1_12 )

[0099] When the frequency fluctuation range is ±0.1Hz, it is considered that the frequency situation meets the quality requirements. The frequency quality index is defined as follows:

[0100]

[0101] Where: T is the total running time; T 0.1 This is the time when the frequency deviation exceeds ±0.1Hz.

[0102] 13) Load forecast accuracy (x 1_13 )

[0103] By formulating relevant assessment standards for load forecasting accuracy, it is helpful to promote load forecasting departments at all levels to improve forecasting accuracy, reduce forecasting errors, and ensure the safe and economical operation of the power grid. Taking the daily load forecasting accuracy as the assessment indicator:

[0104]

[0105]

[0106] Where: F i , T i are the predicted value and actual value of the load at the statistical point, E i is the relative prediction error of the single-point load; i = 1, 2, 3, ..., n; n is the total number of points for daily load prediction. For power grids with daily power loads above 1GW, A I ≥94% is qualified; for power grids with daily power load less than 1GW, A I ≥93% is qualified.

[0107] 14) Average electricity purchase cost deviation rate (x 2_1 )

[0108] In order to accurately reflect the space that can be used to reduce the average power purchase cost through dispatching decisions, the actual power grid parameters of the previous day are used as input, and the balance constraints, safety constraints, fairness constraints and unit performance constraints are considered to obtain the optimal value of the average power purchase cost. Therefore, by setting up the average power purchase cost deviation rate indicator, objective factors such as coal price fluctuations are excluded, and the room for improvement of the average power purchase cost in actual dispatching is examined.

[0109]

[0110] Where: C rea represents the average electricity purchase cost in the actual operation of the power grid on the previous day; C per Represents the average electricity purchasing cost of the power grid under ideal dispatch.

[0111] 15) Network loss rate (x 2_2 )

[0112] The power grid energy rate (line loss rate) is an important economic indicator for the state to assess the power sector. It is also a comprehensive technical and economic indicator that characterizes the planning and design level, production technology level, and operation and management level of the power system.

[0113] Network loss rate = [(input power - output power) / input power] × 100% (18)

[0114] 16) Excess spinning reserve ratio (x) 2_3 )

[0115] The spinning reserve excess rate is used to evaluate the phenomenon of excessive spinning reserve capacity, as follows:

[0116] Spinning reserve excess ratio = max(positive reserve excess ratio, negative reserve excess ratio) (19)

[0117]

[0118] 17) Average load factor of peak load units (x 2_4 )

[0119] By examining the average load rate of peak load units, the rationality of the start-up and shutdown plan can be reflected. If the average load rate of peak load units is low, it means that there are too many units started on that day and the unit utilization rate is low.

[0120]

[0121] Where: P f is the load of the entire network during peak load; C i is the maximum output of unit i; n G For the started units within the province.

[0122] 18) Reactive power balance degree of supply area (x 2_5 )

[0123] Reactive power balance should be achieved as much as possible by stratification (referring to power grids of different voltage levels) and local balancing of different regions to reduce the transmission of reactive power over long distances and between different voltage levels. The high-voltage side of the main transformer of the 220kV substation is taken as an independent assessment unit, and an assessment checkpoint is set on the high-voltage side of the main transformer of the 220kV substation, with a total of n tc donor areas, of which nTj If there are substations, the reactive power balance index of the supply area is as follows:

[0124]

[0125] Where: N Φi Indicates the number of statistical points where the power factor of main transformer i is qualified.

[0126] 19) Average coal consumption for power generation (x 3_1 )

[0127] The average coal consumption for power generation refers to the amount of standard coal consumed by a power generation enterprise for every kilowatt-hour of electricity generated. It is the main indicator for assessing the energy utilization efficiency of power generation enterprises.

[0128] Average coal consumption for power generation (g / kWh) = standard coal consumption for power generation on the previous day / power generation on that day (23)

[0129] The data of standard coal consumption can be collected by a real-time online monitoring system or the actual coal consumption can be approximately simulated with the designed unit coal consumption.

[0130] 20) Equivalent average load factor of generator set (x 3_2 )

[0131] When the generating units are operating in the high load rate range, the coal consumption is relatively low. During the dispatching and operation process, the operating load rate of the started units can be increased by orderly adjusting the coal-fired units and replacing small ones with large ones, thereby obtaining the system coal-saving benefits.

[0132]

[0133] Where: n G is the number of units in operation within the province; M i is the load factor of unit i; a Gi is the unit weight coefficient.

[0134] (2) Building a multi-task time series learning model

[0135] The present invention proposes a multi-task time series learning model based on a long short-term memory network (Long Short-Term Memory, LSTM), which takes the system historical performance score and various system state feature values ​​and their time series as input, and adopts the LSTM deep neural network model to realize the tasks of system performance prediction and key state feature mining. The implementation process is: a multi-level discrete wavelet transform is used to decompose the target time series into a group of subsequences, and the subsequences are fed back to the Seq2Seq predictor, which forms a multi-step prediction of the performance score according to the characteristics of the subsequences. The correlation between the system historical state characteristics and the system historical scores is quantified and fed back to the stacked LSTM module to obtain their time dependency. The Seq2Seq predictor and the hidden layer after the LSTM are superimposed are combined, and the two tasks can share their respective feature representations, be trained together, and be sorted according to the predicted correlation coefficients, so as to identify the key state features.

[0136] The multi-task time series learning model consists of the following five modules:

[0137] 1) Multi-level discrete wavelet transform

[0138] The system performance score is a comprehensive indicator of the information system, which can reflect many aspects of the system. The time series of the system performance score is volatile, contains information of different resolutions in the time / frequency domain, and is difficult to predict directly. Wavelet decomposition is a method for capturing the characteristics of time series in the time domain and the frequency domain. In the present invention, we use a multi-level discrete wavelet transform to decompose the system performance score time series into multiple sub-series arranged from high to low frequency. The decomposed sub-series are more predictable, and their results can be combined to form a more accurate prediction of the original time series.

[0139] 2) Seq2Seq Predictor

[0140] Seq2Seq is a deep model based on recurrent neural network (RNN). The present invention constructs a Seq2Seq model for multi-step prediction. The encoder uses a chain of LSTM units to capture the time-related features of historical observations. The predictor is also a LSTM-based neural network, which uses the encoded vector and the previous time series value as input and outputs multi-step prediction results through internal iterative learning.

[0141] 3) Relevance

[0142] In order to mine the key state characteristics in the system, it is necessary to quantify the correlation coefficient between the system performance score and various states in the system. The present invention adopts the Pearson product-moment correlation coefficient (PCC) method to obtain the correlation between the two.

[0143] 4) Stacked LSTM Model

[0144] LSTM can capture time dependency. In the present invention, a stacked LSTM model with two layers is constructed to capture the multivariate time correlation of high-dimensional time series. The input of the lower LSTM is the correlation coefficient between the system performance score and each state of the system. The output of the lower layer is connected to the input of the upper LSTM. The output of the upper layer is a column of feature vectors. Combined with the input of the Seq2Seq predictor, the correlation coefficient between the future system historical score and the characteristics of each state of the system is predicted.

[0145] 5) Multi-task learning

[0146] Predicting future system historical scores and mining system key state features are two tasks of the present invention, which are connected through a shared hidden layer and share their feature representations with each other. Specifically, the hidden layer of the Seq2Seq predictor and the output of the stacked LSTM model are combined as the input of the full neural network, which outputs the correlation coefficient prediction values ​​of all key state features. According to the ranking of the correlation coefficient, the top K key state features are selected as the output of the model. Since the proposed model combines the hidden layers of the two learning tasks to form the feature representation, the entire network can be jointly trained in a multi-task learning model.

[0147] (3) Self-learning and self-refining of key grid status characteristics

[0148] There is a large amount of time series data in the power grid data, that is, different data have strong correlation in the time dimension. Mining the correlation between time series can significantly improve the accuracy of tasks such as prediction and fault diagnosis. The present invention is based on a multi-task time series learning method, which takes the overall historical operation evaluation value of the system and the characteristic value of the power grid state as input, predicts the overall operation state of the system and mines the key state characteristics in the system.

[0149] The specific steps are as follows:

[0150] The first step is to decompose the input historical time series into multiple subsequences through a multi-level discrete wavelet transform module, the purpose of which is to address the frequency characteristics of the input time series and decompose it into more predictable subsequences.

[0151] Step 2: Feed the decomposed subsequences to the Seq2Seq prediction module. The Seq2Seq predictor is an LSTM-based encoder-decoder that can form multi-step predictions of subsequences individually, and finally integrate the results to output the predicted value of the system state.

[0152] Step 3: Apply the relevance ranking module to quantify the correlation between different grid state characteristics and the system state score, thereby forming a time series of the correlation coefficients of each grid state characteristic in different time periods.

[0153] In the fourth step, the time series of all grid state features are fed into a stacked LSTM module, which uses a two-layer LSTM structure to capture temporal dependencies.

[0154] Step 5: Taking full advantage of the grid state features and time correlation, the Seq2Seq predictor and the hidden layer of the stacked LSTM are combined into the input of a fully connected neural network that predicts the correlation coefficients of all key state features. The predicted correlation coefficients are used for sorting to obtain the top K grid key state features.

[0155] Based on the traditional dispatch decision-making method, the present invention adopts IT technologies such as big data, machine learning, and data mining to establish the intrinsic correlation between the control behavior and the operation status of the power grid, guide and help the control personnel to actively, quickly, comprehensively, and accurately control the current power grid status and development trend, and provide corresponding auxiliary decision-making for the control operation. Fully exploring and giving full play to the value of the large amount of real-time and historical dispatch operation data owned by the power grid is an important method and means to improve the intelligence of the power grid control business.

[0156] Based on the same inventive concept, the present invention also provides a system for determining key state characteristics of a power grid, comprising:

[0157] A calculation module, used for calculating the state characteristic value in the preset characteristic set based on the acquired power grid operation data;

[0158] A result module is used to train a multi-task time series learning model based on all state feature values ​​in the feature set to obtain a predicted value of power grid performance and key state features corresponding to the predicted value of power grid performance;

[0159] Among them, the multi-task time series learning model is constructed based on the relationship between state features and power grid performance; the state features in the feature set are extracted based on power grid security, economy and energy saving.

[0160] In the embodiment, the result module includes:

[0161] A scoring unit, used to obtain the historical performance score of the power grid in time series;

[0162] A prediction unit is used to input the time series and the historical performance score of the power grid corresponding to the time series and the state characteristic value into the multi-level discrete wavelet transform module and the Seq2Seq predictor in sequence to obtain the predicted value of the power grid performance;

[0163] A coefficient unit, used to input the time series and the historical performance score of the power grid corresponding to the time series and the state characteristic value into the correlation module and the Seq2Seq predictor in sequence to obtain the correlation coefficients of all state characteristics;

[0164] A result unit, used to sort the correlation coefficients of all state characteristics, and select the key state characteristics of the power grid corresponding to the predicted value of the power grid performance in order;

[0165] The Seq2Seq predictor includes multiple LSTM encoder-decoders and hidden layers.

[0166] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.

[0167] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0168] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0169] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0170] The above are merely embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention are included in the scope of the claims of the present invention to be approved.

Claims

1. A method for determining key state characteristics of a power grid, characterized in that: include: Calculating state characteristic values ​​in a preset characteristic set based on the acquired power grid operation data; Based on all state feature values ​​in the feature set, a multi-task time series learning model is used for training to obtain a predicted value of power grid performance and key state features corresponding to the predicted value of power grid performance; The multi-task time series learning model is constructed based on the relationship between state features and power grid performance; the state features in the feature set are extracted based on power grid security, economy and energy saving; The method of training all state feature values ​​in the feature set using a multi-task time series learning model to obtain a predicted value of power grid performance and key state features corresponding to the predicted value of power grid performance includes: Obtain historical performance scores of power grid in chronological order; The time series and the historical performance scores of the power grid corresponding to the time series and the state characteristic values ​​are sequentially input into the multi-level discrete wavelet transform module and the Seq2Seq predictor to obtain the predicted value of the power grid performance; Input the time series and the historical performance scores of the power grid corresponding to the time series and the state characteristic values ​​into the correlation module and the Seq2Seq predictor in sequence to obtain the correlation coefficients of all state characteristics; Sorting the correlation coefficients of all state characteristics, and selecting the key state characteristics of the power grid corresponding to the predicted values ​​of the power grid performance in order; The Seq2Seq predictor includes multiple LSTM encoder-decoders and hidden layers; The time series and the power grid historical performance score and state characteristic value corresponding to the time series are sequentially input into the multi-level discrete wavelet transform module and the Seq2Seq predictor to obtain the predicted value of the power grid performance, including: The time series and the historical performance scores of the power grid corresponding to the time series and the state characteristic values ​​are input into a multi-level discrete wavelet transform module to decompose the input time series into multiple subsequences; Feed the decomposed subsequences to the Seq2Seq predictor to output the predicted value of power grid performance; The time series and the historical performance score of the power grid corresponding to the time series and the state characteristic value are sequentially input into the correlation module and the Seq2Seq predictor to obtain the correlation coefficients of all state characteristics, including: Input the time series and the historical performance scores of the power grid corresponding to the time series and the state characteristic values ​​into the correlation module; Obtaining a time series of correlation coefficients of each state feature in different time periods based on the correlation module; Feed the time series of all state features into the stacked LSTM module; The output of the stacked LSTM module is input into the hidden layer of the Seq2Seq predictor, and the output of the hidden layer is input into a fully connected neural network to obtain the correlation coefficients of all state features.

2. The method according to claim 1, characterized in that The power grid historical performance score is obtained through expert scoring or set weights.

3. The method according to claim 1, characterized in that The state features extracted based on security include: N-1 pass rate, N-2 pass rate of main sections, short-circuit current, power flow operation, main transformer safety, line safety, grid operation balance, rotating reserve deficiency rate, supply area load reactive power margin, voltage safety, frequency qualification rate, frequency quality rate and load forecast accuracy.

4. The method according to claim 1, characterized in that The state characteristics for economic extraction include: The average electricity purchase cost deviation rate, network loss rate, rotating reserve excess rate, average load rate of peak load units and reactive power balance degree of the supply area.

5. The method according to claim 1, characterized in that The state features extracted based on energy efficiency include: Average coal consumption for power generation and equivalent average load factor of generating units.

6. A system for determining key state characteristics of a power grid, characterized in that: include: A calculation module, used for calculating the state characteristic value in the preset characteristic set based on the acquired power grid operation data; A result module is used to train a multi-task time series learning model based on all state feature values ​​in the feature set to obtain a predicted value of power grid performance and key state features corresponding to the predicted value of power grid performance; The multi-task time series learning model is constructed based on the relationship between state features and power grid performance; the state features in the feature set are extracted based on power grid security, economy and energy saving; The result module includes: A scoring unit, used to obtain the historical performance score of the power grid in time series; A prediction unit is used to input the time series and the historical performance score of the power grid corresponding to the time series and the state characteristic value into the multi-level discrete wavelet transform module and the Seq2Seq predictor in sequence to obtain the predicted value of the power grid performance; A coefficient unit, used to input the time series and the historical performance score of the power grid corresponding to the time series and the state characteristic value into the correlation module and the Seq2Seq predictor in sequence to obtain the correlation coefficients of all state characteristics; A result unit, used to sort the correlation coefficients of all state characteristics, and select the key state characteristics of the power grid corresponding to the predicted value of the power grid performance in order; The Seq2Seq predictor includes multiple LSTM encoder-decoders and hidden layers; The prediction unit is specifically used for: The time series and the historical performance scores of the power grid corresponding to the time series and the state characteristic values ​​are input into a multi-level discrete wavelet transform module to decompose the input time series into multiple subsequences; Feed the decomposed subsequences to the Seq2Seq predictor to output the predicted value of power grid performance; The coefficient unit is specifically used for: Input the time series and the historical performance scores of the power grid corresponding to the time series and the state characteristic values ​​into the correlation module; Obtaining a time series of correlation coefficients of each state feature in different time periods based on the correlation module; Feed the time series of all state features into the stacked LSTM module; The output of the stacked LSTM module is input into the hidden layer of the Seq2Seq predictor, and the output of the hidden layer is input into a fully connected neural network to obtain the correlation coefficients of all state features.

Images