Wind power plant reactive power regulation and control method based on source load characteristic data analysis

Abstract

The invention discloses a wind power plant reactive power regulation and control method based on source load characteristic data analysis, and relates to the technical field of power regulation and control, and the method comprises the steps: presetting a demand end sliding window and a power supply end sliding window; carrying out source load characteristic data acquisition to obtain demand end characteristic data and power supply end characteristic data; carrying out regional distributed reactive power demand prediction to obtain a real-time reactive power demand; performing wind power plant partition reactive power prediction to obtain real-time reactive power data; obtaining a reactive power distribution strategy; and dynamically regulating and controlling the reactive power of the wind power plant. The technical problems that reactive power regulation and control are not accurate, distribution and utilization are not reasonable and stable and efficient operation is difficult to guarantee due to varied source load characteristics of the wind power plant in the prior art are solved, dynamic and accurate regulation and control of the reactive power of the wind power plant according to the source load characteristics are achieved, reactive power distribution and utilization are improved, and the wind power generation efficiency is improved. And stable and efficient operation of the wind power plant is ensured.

Description

Technical Field

[0001] The present invention relates to the technical field of power regulation, and in particular to a method for regulating reactive power of a wind farm by analyzing source-load characteristic data. Background Art

[0002] During the operation of traditional wind farms, the output power of wind farms is extremely unstable due to the natural intermittent and volatile nature of wind energy. This directly leads to frequent and large fluctuations in the reactive power connected to the power grid, and the early wind farm monitoring and control systems often lack detailed and real-time insight into the source and load ends. On the one hand, it is impossible to accurately capture the dynamic demand for reactive power of power loads in different time periods and different regions, resulting in large deviations in reactive power prediction on the demand side; on the other hand, there is also a lack of effective prediction methods for the changes in reactive power generated by power generation in each sub-area of ​​the wind farm, resulting in unclear characteristics of the power supply end. In addition, in the reactive power distribution link, only extensive strategies can be adopted, and it is impossible to flexibly allocate according to the real-time changing source and load characteristics, which not only causes a waste of reactive power, but also easily causes problems such as voltage fluctuations and reduced stability of the power grid, seriously affecting the stable and efficient operation of wind farms, and it is difficult to meet the growing demand for power supply and grid stability requirements.

[0003] The existing technology has technical problems such as inaccurate reactive power control, unreasonable distribution and utilization due to variable source-load characteristics of wind farms, making it difficult to ensure stable and efficient operation. Summary of the invention

[0004] The present application provides a wind farm reactive power control method based on source-load characteristic data analysis, which is used to solve the technical problems in the prior art that the reactive power control of wind farms is inaccurate, the allocation and utilization are unreasonable, and it is difficult to ensure stable and efficient operation due to the variable source-load characteristics.

[0005] In view of the above problems, the present application provides a wind farm reactive power control method based on source-load characteristic data analysis, the method comprising:

[0006] A demand-side sliding window and a power-side sliding window are preset; the demand-side sliding window and the power-side sliding window are used as data collection time limits to collect source-load characteristic data and obtain demand-side characteristic data and power-side characteristic data; regional distributed reactive power demand forecasting is performed on the demand-side characteristic data to obtain real-time reactive power demand; wind farm partition reactive power forecasting is performed based on the power-side characteristic data to obtain real-time reactive power data; reactive power cloud receives and optimizes reactive power partition control according to the real-time reactive power demand and real-time reactive power data to obtain a reactive power allocation strategy; the reactive power allocation strategy is loaded into the reactive regulation device and SVG device of the target wind farm, and the reactive power of the wind farm is dynamically regulated through the reactive regulation device and SVG device.

[0007] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0008] Preset the demand-side sliding window and the power supply-side sliding window; collect source-load characteristic data to obtain demand-side characteristic data and power supply-side characteristic data; predict regional distributed reactive power demand based on the demand-side characteristic data to obtain real-time reactive power demand; predict wind farm zoning reactive power based on the power supply-side characteristic data to obtain real-time reactive power data; optimize reactive power zoning control to obtain reactive power allocation strategy; and dynamically control wind farm reactive power. The technical effect of achieving dynamic and precise control of reactive power based on source-load characteristics, improving reactive power allocation and utilization, and ensuring stable and efficient operation of wind farms has been achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0010] Figure 1 A schematic flow chart of a wind farm reactive power control method based on source-load characteristic data analysis provided in an embodiment of the present application;

[0011] Figure 2 A schematic diagram of a process for obtaining real-time reactive power demand in a wind farm reactive power control method based on source-load characteristic data analysis provided in an embodiment of the present application. DETAILED DESCRIPTION

[0012] The present application provides a wind farm reactive power control method based on source-load characteristic data analysis, which is used to solve the technical problems in the prior art that wind farms have inaccurate reactive power control, unreasonable allocation and utilization, and difficulty in ensuring stable and efficient operation due to variable source-load characteristics.

[0013] The technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0014] Examples, such as Figure 1 As shown, the present application provides a wind farm reactive power control method based on source-load characteristic data analysis, the method comprising:

[0015] Step S100: Preset a demand-side sliding window and a power-side sliding window.

[0016] Specifically, in the initial stage of the reactive power control process of the wind farm in the source-load characteristic data analysis, the demand-side sliding window and the power supply-side sliding window are precisely preset. The preset of the demand-side sliding window is based on the in-depth analysis of the power load characteristics and the data collection accuracy requirements. Through in-depth mining and statistical analysis of historical power consumption data, its initial time point is accurately determined to coincide with the starting time of the typical power consumption cycle, so as to fully cover the complete change process of the power load. At the same time, according to the dynamic change frequency of power demand and data processing capabilities, the time span range is scientifically set to ensure that the fluctuation characteristics of power load in different periods can be effectively captured. In addition, the sliding step size is strictly determined. The step size is closely related to the time resolution of the required data. Through precise calculation and optimization, it is ensured that sufficiently fine and representative data samples can be obtained when collecting demand-side characteristic data, thereby accurately reflecting the real-time change trend of power demand.

[0017] Similarly, the preset sliding window of the power supply end is also carefully designed based on the unique nature of power supply. Taking into account the output characteristics of various power sources (such as wind power generation affected by wind speed and wind direction changes, photovoltaic power generation affected by light intensity and duration fluctuations, etc.) and the complex matching relationship with electricity demand, the initial time, time range and sliding step are accurately defined. Through a detailed analysis of local long-term meteorological data and combined with the operating characteristics of wind power generation equipment, the initial time of the sliding window is set at a time that is compatible with the start-up characteristics of wind power generation. The time range is dynamically adjusted according to the seasonal and daily changes of wind power to ensure that the power generation output under different wind conditions can be fully covered. The sliding step is optimized and determined based on the stability of power generation data and analysis requirements, so that the key information of power supply can be accurately obtained when collecting power supply end characteristic data, providing a solid data foundation for subsequent reactive power prediction and regulation. These two sliding windows have built a solid time frame cornerstone for the precise collection of subsequent source-load characteristic data, and achieved precise control of wind farm reactive power.

[0018] Step S200: Using the demand-side sliding window and the power-side sliding window as data collection time limits, collecting source-load characteristic data, and obtaining demand-side characteristic data and power-side characteristic data.

[0019] Specifically, the preset demand-side sliding window and power-side sliding window are used as precise data collection time limits to fully start the source-load characteristic data collection work. During the data collection process, for the demand side, power consumption information is obtained from power facilities in different regions, covering the overall power consumption trend of the region, load change rules, etc., so as to summarize the demand-side characteristic data, which reflects the demand for reactive power in different regions within a specific time.

[0020] At the same time, for the power supply end, with the power supply end sliding window as the time constraint, all-round data collection is carried out on the power generation equipment in the wind farm. Sensors installed at various key parts of the wind turbine are used to obtain key operating data such as wind speed, wind direction, generator speed, and output power in real time. At the same time, combined with the voltage, current and other parameters of the grid access point, the operating status of the power generation equipment is fully understood. According to the time rules of the power supply end sliding window, the collected data is organized and processed in an orderly manner, and the power supply end characteristic data reflecting the power supply characteristics are extracted from the massive data, including the power generation power curve under different wind speed conditions, reactive power output characteristics, unit operation stability indicators, etc. Through this refined data collection and processing method based on the sliding window, it is ensured that the acquired demand-side characteristic data and power supply-side characteristic data have a high degree of timeliness and accuracy, providing solid data support for subsequent reactive power demand forecasting and zoning regulation.

[0021] Step S300: performing regional distributed reactive power demand forecasting on the demand-side characteristic data to obtain real-time reactive power demand.

[0022] Specifically, by interacting with the power grid management system, the power supply coverage information of the target wind farm is accurately obtained. Based on this, a complex topology analysis algorithm is used to screen and generate the target power grid topology structure. After accurately locating the substation in the topology, the target power grid topology is divided into multiple distributed power supply areas according to the current flow direction using a graph theory algorithm. Subsequently, the rich partitioned power consumption characteristic data of each distributed power supply area is called from the local database, including key indicators such as the total regional load, the proportion of regional load types, the standard deviation of regional load, the regional load change rate, and the regional load fluctuation ratio. Based on these partitioned power consumption characteristics, regional clustering analysis is carried out. First, multiple distributed power supply areas are permuted and enumerated. For each group of power supply areas, the Euclidean distance of their partitioned power consumption characteristics is calculated using a mathematical model to obtain a large number of regional similarity coefficients. Each distributed power supply area is regarded as a topological node, and the regional similarity coefficient is used as the length of the topological connection. Based on this, a regional similarity feature topology is constructed. A regional similarity threshold is preset, and the topological connection is deleted according to the threshold. The regional similarity feature topology is split into multiple local similarity feature topologies, and its topological nodes are extracted to obtain M groups of distributed power supply areas. For these M groups of distributed power supply areas, dynamic reactive power prediction models are constructed respectively. Taking the first group of distributed power supply areas as an example, the historical reactive power demand and historical external meteorological characteristics are called, and the historical reactive power demand is finely decomposed using the demand-side sliding window to obtain multiple groups of stage reactive power demand. The stage reactive power sequence is generated by summing the intra-group time series mapping; at the same time, the historical external meteorological characteristics are decomposed to obtain the stage external meteorological sequence and the time feature coding mark is assigned. The time window closing point of the stage reactive power sequence is collected, and the prediction time window is set. The stage reactive power sequence is segmented as the starting point of data extraction to obtain the updated reactive power sequence. The stage external meteorological sequence, stage time feature coding and stage reactive power sequence are used as training data. A time series feature extraction unit is constructed based on the CNN model, and a time series power prediction unit is constructed based on the LSTM model. The two are connected to complete the construction of the first dynamic reactive power prediction model and jointly trained. M dynamic reactive power prediction models are constructed in this way, and finally the reactive power prediction collaborative model is constructed by connecting these M models in parallel. The demand-side characteristic data is input into it. After complex calculation and model reasoning, the real-time reactive power demand is finally output, providing accurate reactive power demand basis for reactive power regulation of wind farms.

[0023] Step S400: performing reactive power prediction of wind farm partitions based on the power supply end characteristic data to obtain real-time reactive power data.

[0024] Specifically, the reactive power prediction of wind farm partitions is carried out using the characteristic data of the power supply end to obtain real-time reactive power data, interact with related systems, and accurately obtain the detailed unit power generation characteristics of P wind turbines in the target wind farm. These characteristics cover key information such as the power generation efficiency and power output curve of the units under different wind speeds and wind directions. Based on these unit power generation characteristics, the topological analysis algorithm is used to deeply analyze the connection relationship, electrical characteristics, and geographical distribution of the P wind turbines. According to scientific partitioning rules, they are divided into L wind farm partitions to ensure that the wind turbines in each partition have certain similarities and correlations in operating characteristics and electrical connections. After the partitioning is completed, the historical data call work is carried out comprehensively for the L wind farm partitions, and the corresponding L groups of historical power generation data are accurately extracted from the massive database. These data contain the change records of the power generation power, active power, reactive power, and related operating parameters of the wind turbines in each partition in different time periods in the past. Then, the 1 / y sliding window of the power supply end is used as a fine data segmentation scale to deeply decompose the L groups of historical power generation data, and the professional data processing algorithm is used to accurately separate the L groups of historical reactive power sequences from the power generation data. These sequences clearly present the changing trend and fluctuation law of reactive power in each partition on the historical time axis. At the same time, according to the terminal time node of the L group of historical reactive power sequences and the time range of the power supply end sliding window, the wind power environment data collection program is started. By deploying meteorological sensors at various key locations of the wind farm and docking with the external meteorological data platform, the multi-dimensional wind power environment data such as wind speed, wind direction, temperature, and air pressure related to the time period are comprehensively collected and integrated into L groups of wind power environment data sequences, providing rich environmental information input for subsequent prediction models. Subsequently, the LSTM model is used to pre-build a standard prediction model, clarifying that the input data of the model is the wind power environment data sequence, the output data is the reactive power data sequence, and the time ratio of the output result to the input result is strictly set to 1 / y to ensure the accuracy and consistency of the prediction results in the time dimension. L groups of wind power environment data sequences and L groups of historical reactive power sequences are used as training data, which are input into the standard prediction model. Advanced machine learning algorithms are used to update the model parameters in a directional manner. Through multiple iterations of training, the weight coefficients and neuron connection relationships within the model are continuously optimized until the model achieves a high prediction accuracy on the training data. After the training is completed, L reactive power prediction units are obtained, each of which contains L groups of reactive power prediction branches in parallel, and the prediction efficiency and accuracy are further improved through parallel computing. Finally, the L reactive power prediction units are connected in parallel to successfully construct a reactive power prediction model.The power supply end characteristic data is fully input into the prediction model. Based on the pre-trained parameters and algorithm logic, the model makes an in-depth prediction of the reactive power changes of each wind farm partition in the current and future period. After a complex calculation process, it finally outputs real-time reactive power data. These data accurately reflect the real-time status of reactive power of each partition of the wind farm under different operating conditions, and provide key data support for the subsequent dynamic regulation of reactive power.

[0025] Step S500: The reactive power flow cloud receives and optimizes reactive power partitioning control according to the real-time reactive power demand and real-time reactive power data to obtain a reactive power allocation strategy.

[0026] Specifically, as the core control unit, reactive flow cloud first receives the real-time reactive power demand and real-time reactive power data accurately calculated from the previous steps. Then, it conducts in-depth decomposition operations on the reactive power data of L groups in the real-time reactive power data, and uses complex algorithms and data analysis models to accurately extract the upper limit of the single-machine output of each of the P wind turbines. This upper limit reflects the reactive power output limit that each wind turbine may reach under the current operating state. At the same time, by interacting with the wind turbine control system, the lower limit of the single-machine output of the P wind turbines is obtained, which clarifies the minimum reactive power output value required for the stable operation of the wind turbine. Based on the obtained P single-machine output upper limits and P single-machine output lower limits, reactive power regulation constraints are constructed and used as boundary limits for the subsequent optimization strategy generation. Under these constraints, the P single-machine output upper limits are finely updated, and the optimization algorithm and strategy adjustment mechanism are used to comprehensively consider the operating efficiency, stability, and matching degree of the wind turbine with the real-time reactive power demand, and obtain P updated single-machine output strategies. Based on the real-time reactive power demand and the newly generated P updated single-machine output strategies, a professional power system analysis algorithm is used to generate the first SVG regulation strategy, which clarifies the reactive power regulation direction and amplitude of the SVG equipment under the current working conditions.

[0027] Then, a comprehensive reactive power control simulation is performed using P updated single-machine output strategies and the first SVG regulation strategy. During the simulation, an accurate power system simulation model is used to fully consider the electrical characteristics of each device in the wind farm, the grid connection status, and the mutual influence under different operating conditions, and the SVG operation loss coefficient is calculated. This coefficient accurately reflects the energy loss of the SVG device when executing the regulation strategy; at the same time, the grid voltage stability is obtained. This indicator intuitively represents the degree of influence of the regulation strategy on the grid voltage stability; and the L partition stability coefficients of the L wind farm partitions. These coefficients evaluate the influence of the regulation strategy on the internal stability of the wind farm from the perspective of different partitions. Based on the above calculation results, a pre-set weighted evaluation system is used to conduct a comprehensive weighted evaluation of the SVG operation loss coefficient, the grid voltage stability, and the L partition stability coefficients, and the first regulation stability is output. This stability index comprehensively measures the overall performance of the current reactive power regulation strategy. Then, according to the operation requirements and stability standards of the wind farm, a regulation stability threshold is preset, which serves as the key basis for judging whether the regulation strategy meets the standards. By calculating the deviation percentage between the control stability threshold and the first control stability, the feedback control algorithm is used to fine-tune the P updated single-machine output strategies and the first SVG control strategy based on this deviation information, and the control stability is evaluated again based on the adjusted strategy. This is repeated iteration, and the strategy combination is continuously optimized until the reactive power allocation strategy is obtained whose control stability data meets the preset control stability threshold. This strategy ultimately determines the optimal configuration of the P single-machine power output strategies and the SVG power control strategy, providing a scientific and effective guidance scheme for the precise dynamic control of reactive power in wind farms.

[0028] Step S600: loading the reactive power allocation strategy into the reactive power regulation device and SVG device of the target wind farm, and dynamically regulating the reactive power of the wind farm through the reactive power regulation device and SVG device in collaboration.

[0029] Specifically, the carefully formulated reactive power allocation strategy is accurately loaded into the reactive power regulation equipment and SVG equipment of the target wind farm. The P single-machine power output strategies and SVG power regulation strategies contained in the reactive power allocation strategy are transmitted to the corresponding equipment in a high-speed and stable manner through a dedicated data transmission channel. The reactive power regulation equipment quickly adjusts its internal control parameters based on the received single-machine power output strategy, such as adjusting the tap position of the transformer, changing the number of switching groups of the reactor, etc., to achieve precise control of the reactive power output of the wind turbine. The SVG equipment uses its advanced power electronic conversion technology to quickly adjust the output reactive power size and phase according to the SVG power regulation strategy to flexibly respond to changes in the system's reactive power demand. During the regulation process, the reactive power regulation equipment and the SVG equipment maintain close communication and collaboration, and form an organic collaborative regulation system by sharing information such as equipment operating status and reactive power output in real time. When the operating conditions of a wind farm change, such as sudden changes in wind speed or load fluctuations, the reactive power regulation equipment and SVG equipment can dynamically adjust their reactive power outputs based on real-time monitoring data and reactive power allocation strategies to ensure that the reactive power of the wind farm is always maintained within a reasonable range, ensuring grid voltage stability and power factor compliance, and achieving efficient and dynamic regulation of wind farm reactive power, thereby improving the operating stability and reliability of the entire wind farm and grid system.

[0030] In one possible implementation, Figure 2 As shown, step S300 also includes:

[0031] Step S310: Interactively obtain the power supply coverage information of the target wind farm, and filter and generate the target power grid topology based on the power supply coverage information. Step S320: After locating the substation in the target power grid topology, divide the target power grid topology into multiple distributed power supply areas according to the current flow direction. Step S330: Locally call to obtain the power consumption characteristics of multiple partitions of the multiple distributed power supply areas. Step S340: Perform regional clustering analysis based on the power consumption characteristics of multiple partitions of the multiple distributed power supply areas to obtain M groups of distributed power supply areas. Step S350: Construct M dynamic reactive power prediction models for the M groups of distributed power supply areas, and complete the construction of the reactive power prediction collaborative model by connecting the M dynamic reactive power prediction models in parallel. Step S360: The reactive power prediction collaborative model uses the demand-side characteristic data as input to predict the regional distributed reactive power demand, and outputs the real-time reactive power demand.

[0032] Specifically, by interacting with external platforms such as the power grid dispatching system and geographic information system, the target wind farm’s accurate power supply coverage information is obtained, which covers key elements such as the geographical scope of the wind farm’s power supply area, the distribution of electricity users, and the grid access point. Based on this rich power supply coverage information, the topology analysis algorithm and the power system knowledge base are used to screen and generate a target grid topology that is highly consistent with the actual grid structure, clearly presenting the connection relationship between each node, line, and equipment in the grid.

[0033] According to the unique electrical characteristics and identification information of the substation in the target power grid topology, the node identification algorithm is used to accurately locate each substation node. As a key power supply hub, these substations play a core role in power conversion and distribution in the entire power grid structure. After successfully locating the substation, the electrical parameters and connection relationships of each line in the power grid topology are deeply analyzed, and the flow path of the current in the power grid is accurately tracked on this basis. Starting from the substation as the power supply source, along the direction of the current flow, the influence of factors such as line resistance and reactance on the current distribution is comprehensively considered. The target power grid topology is carefully divided into multiple distributed power supply areas using the regional partitioning algorithm in graph theory and the power system flow analysis theory. Each distributed power supply area takes a substation as the core power supply center. All user nodes directly powered by the substation constitute the power consumption range of this area. In the division process, it is ensured that each distributed power supply area is relatively independent in electricity, and at the same time, it can be interconnected with other areas through high-voltage transmission lines to ensure the stability and reliability of the entire power grid system. This division method based on current flow direction and substation location makes the power consumption characteristics and reactive power demand in each distributed power supply area have a strong correlation and consistency, providing a reasonable spatial architecture foundation for the subsequent accurate analysis of regional power consumption characteristics and the construction of reactive power prediction models, which helps to improve the pertinence and effectiveness of reactive power regulation, thereby better meeting the stable operation needs of wind farms and the entire power system.

[0034] Establish a stable data connection channel with the local power data storage center, and accurately locate and extract massive historical power consumption data related to multiple distributed power supply areas through carefully designed data query instructions. These data cover rich information such as operating parameters, power consumption records, and power consumption time distribution of all power equipment in each area in different time periods. Calculate the total regional load of each distributed power supply area at different time scales to accurately grasp the scale of power consumption in each area; then analyze the proportion of various types of loads in the region (such as residential electricity, industrial production electricity, commercial electricity, etc.) to clarify the load type structure; at the same time, calculate the regional load standard deviation to evaluate the discrete degree of power load and reflect the stability of load fluctuations; further analyze the regional load change rate to gain insight into the trend of power demand over time; finally calculate the regional load fluctuation ratio to measure the fluctuation amplitude of load in different time periods. Through this series of complex data processing operations, multiple regional power consumption characteristics are extracted, including key indicators such as regional total load, regional load type ratio, regional load standard deviation, regional load change rate and regional load fluctuation ratio. These characteristic data comprehensively and accurately characterize the power consumption behavior patterns and demand change laws of each distributed power supply area, providing a solid data foundation for subsequent regional clustering analysis and reactive power demand forecasting, making the analysis of power demand more in-depth and accurate, thereby providing strong support for the formulation of reactive power control strategies for wind farms.

[0035] Comprehensive local historical electricity consumption data calls are carried out for multiple distributed power supply areas to obtain rich multiple partition electricity consumption characteristic data, which covers key information such as regional total load, regional load type ratio, regional load standard deviation, regional load change rate and regional load fluctuation ratio. Permutation and combination enumeration operations are performed on these distributed power supply areas to exhaust all possible power supply area pairing combinations. For each group of power supply area pairing, the Euclidean distance between its partition electricity consumption characteristics is calculated using a mathematical model. In this way, many regional similarity coefficients are obtained, which accurately quantify the similarity of electricity consumption characteristics between different regions. Then, multiple distributed power supply areas are abstracted as topological nodes, and the calculated regional similarity coefficients are used as the topological connection length. The corresponding topological nodes are connected according to the pairing relationship of each power supply area, so as to construct a regional similarity feature topology. On this basis, a regional similarity threshold is pre-set, which is used as the key criterion for distinguishing regional similarity. By performing the topological connection deletion operation of the regional similarity feature topology based on this threshold, the entire regional similarity feature topology is split into M local similarity feature topologies, and the regions represented by the nodes in each local similarity feature topology have high similarity in electricity consumption characteristics. Finally, the topological node extraction operation is performed on M local similar feature topologies to obtain M groups of distributed power supply areas, and several partitions in each group show similar power consumption characteristics. This regional clustering analysis method can effectively identify regional groups with similar power consumption behavior patterns, laying a solid foundation for the subsequent construction of targeted reactive power prediction models, so that the prediction model can better adapt to the power consumption characteristics of different regional groups, thereby improving the accuracy and effectiveness of reactive power demand prediction, and further optimizing the formulation of wind farm reactive power control strategies.

[0036] A dynamic reactive power prediction model is constructed for each of the M groups of distributed power supply areas. Taking the first group of distributed power supply areas as an example, the historical reactive power demand data and the corresponding historical external meteorological characteristic data are first called from the database, and the historical reactive power demand is decomposed into H groups of reactive power demand according to the time series using the demand-side sliding window. The reactive power demand of each group of stages is summed up by intra-group time series mapping to generate H stages of reactive power sequences; at the same time, the historical external meteorological characteristics are decomposed in the same way to obtain H stages of external meteorological sequences, and each stage is given a time feature coding identifier. The time window closing point of the H stage reactive power sequence is collected, 1 / t of the demand-side sliding window is set as the prediction time window, and the time window closing point is used as the starting point for data extraction. The H stage reactive power sequences are segmented according to the prediction time window to obtain H updated reactive power sequences. Using H phases of external meteorological sequences, H phases of time feature codes, and H phases of reactive power sequences as training data, a time series feature extraction unit is constructed based on the CNN model to extract the spatiotemporal features of the data; a time series power prediction unit is constructed based on the LSTM model to predict the reactive power change trend. The output of the time series feature extraction unit is connected to the input of the time series power prediction unit to complete the construction of the first dynamic reactive power prediction model. Similarly, M dynamic reactive power prediction models are constructed, and the construction of the reactive power prediction collaborative model is completed by connecting these M models in parallel to improve the accuracy and robustness of the prediction.

[0037] The reactive power prediction collaborative model takes the demand-side characteristic data as input, and uses the complex neural network structure and trained parameters built inside it to deeply predict the regional distributed reactive power demand. After a large amount of calculation and model reasoning, it finally outputs the real-time reactive power demand, providing an accurate basis for the reactive power demand on the power consumption side for the reactive power control of the wind farm, ensuring that the control strategy can closely fit the actual power demand and realize the scientific and efficient control of the reactive power of the wind farm.

[0038] In a possible implementation, step S400 further includes:

[0039] Step S410: interactively obtain the power generation characteristics of P wind turbines in the target wind farm. Step S420: perform topological analysis on the P wind turbines based on the power generation characteristics of the P wind turbines, and divide the P wind turbines into L wind farm partitions. Step S430: call historical data for the L wind farm partitions to obtain L groups of historical power generation data. Step S440: use 1 / y of the power supply end sliding window as the data segmentation scale, decompose the L groups of historical power generation data, and obtain L groups of historical reactive power sequences. Step S450: collect wind power environment data according to the L group terminal time nodes of the L groups of historical reactive power sequences and the power supply end sliding window, and obtain L groups of wind power environment data sequences. Step S460: use the LSTM model to pre-build a standard prediction model, wherein the input data of the standard prediction model is a wind power environment data sequence, the output data of the standard prediction model is a reactive power data sequence, and the time ratio of the output result to the input result is 1 / y. Step S470: Use the L groups of wind power environment data sequences and the L groups of historical reactive power sequences as training data to perform directional parameter updates on the standard prediction model, and obtain L reactive power prediction units, wherein the L reactive power prediction units include L groups of reactive power prediction branches connected in parallel. Step S480: Connect the L reactive power prediction units in parallel to complete the construction of the reactive power prediction model. Step S490: Input the power supply end characteristic data into the reactive power prediction model to perform wind farm partition reactive power prediction, and obtain the real-time reactive power data, wherein the real-time reactive power data includes L groups of partition reactive power data.

[0040] Specifically, by interacting with the wind farm monitoring system, the unit management system and related databases, the detailed P wind turbine generation characteristics of the target wind farm are obtained. These characteristics include key parameters such as the power output curve, power generation efficiency, start-up wind speed, and shutdown wind speed of the wind turbine under different wind speed and wind direction conditions, as well as data related to the unit's operating stability and reliability, providing basic information for subsequent analysis.

[0041] Based on the acquired power generation characteristics of P units, a complex topological analysis algorithm is used to comprehensively analyze the P wind turbines. Taking into account the geographical distribution, electrical connection relationship, similarity of power generation characteristics and other factors of the wind turbines, the P wind turbines are scientifically and rationally divided into L wind farm zones. The wind turbines in each zone have high consistency in terms of operating characteristics and response to environmental changes, which facilitates subsequent zone management and prediction.

[0042] For the L wind farm partitions, an accurate data call request is sent to the local historical database to obtain the corresponding L groups of historical power generation data. These data record the detailed operating parameters such as power generation power, active power, reactive power, voltage, current, etc. of each partition in different time periods in the past, as well as the corresponding timestamp information, providing rich data materials for analyzing the power generation rules of the partitions.

[0043] The 1 / y sliding window of the power supply end is set as a fine data segmentation scale, and a professional data processing algorithm is used to deeply decompose the L group of historical power generation data. By cutting and analyzing the power generation data in the time dimension according to this scale, the L group of historical reactive power sequences are accurately extracted, and the change trajectory and fluctuation characteristics of the reactive power of each partition over time are clearly presented, laying the foundation for the subsequent correlation analysis with wind power environment data.

[0044] The time range of data collection is determined based on the terminal time node of the L group of historical reactive power sequences, and the wind power environment data collection program is started in combination with the time constraints of the sliding window at the power supply end. Through the meteorological sensor network deployed in the wind farm and the data connection with the external meteorological data platform, multi-dimensional wind power environment data such as wind speed, wind direction, temperature, air pressure, and air humidity in this time period are collected and integrated into L group wind power environment data sequences, so that the power generation data and environmental data can be accurately matched in time and space.

[0045] Based on the demand for reactive power prediction of wind farms, an LSTM model with strong time series processing capabilities is selected to pre-build a standard prediction model. Due to its unique memory unit structure, the LSTM model can effectively capture the long-term dependencies in the wind power environment data sequence and is very suitable for processing wind power data with time series characteristics. The core of the design of this standard prediction model is to clarify its input and output data types and the time relationship between the two. The input data is a wind power environment data sequence, which contains the changes of multi-dimensional environmental information such as wind speed, wind direction, temperature, and air pressure over time. The duration is a specific time period. These data fully reflect the dynamic change characteristics of the wind farm operating environment and are key external factors affecting the reactive power output of wind turbines. The output data of the model is a reactive power data sequence, and the time ratio of its output result to the input result is set to 1 / y, which means that based on the input wind power environment data sequence with a duration of T, the model will predict the output change trend of reactive power within the future T / y duration, thereby realizing the effective estimation of reactive power output on the power generation side and providing a key basis for subsequent reactive power zoning control. When constructing the model, the network structure of the LSTM model is carefully designed, and key parameters such as the appropriate number of hidden layers, number of neurons, and activation function are determined to ensure that the model has sufficient expression and learning capabilities to accurately describe the complex nonlinear relationship between wind power environment data and reactive power data. At the same time, for the model training process, a suitable loss function, such as the mean square error function, is selected to measure the difference between the predicted results and the actual reactive power data. Through optimization algorithms, such as the stochastic gradient descent algorithm, the model parameters are continuously adjusted to make the model gradually converge during the training process, improve the accuracy and reliability of the prediction, and thus lay a solid model foundation for the subsequent reactive power zoning prediction of wind farms and the effective operation of the entire control system.

[0046] The Adagrad adaptive learning rate optimization algorithm is used to implement the directional parameter update of the standard prediction model, so as to obtain L reactive power prediction units, each of which contains L parallel reactive power prediction branches. Data integration and preprocessing are performed on L wind power environment data sequences and L historical reactive power sequences to construct a training data set. The wind power environment data sequence is used as the input feature and the historical reactive power sequence is used as the target output to ensure the correspondence of the data. Then, the data set is randomly divided into a training set, a validation set, and a test set, with a ratio of 70%, 20%, and 10%, respectively. In the training stage, data samples are extracted from the training set in batches, and the wind power environment data sequence is input into the standard prediction model. Based on the current parameters, the model performs forward propagation calculations through the LSTM network structure, and uses the unique gating mechanism of LSTM to process the time series information to obtain the predicted reactive power data sequence. Then, the mean square error (MSE) between the predicted sequence and the actual historical reactive power sequence is calculated. After that, the Adagrad algorithm is used to back propagate and update the model parameters based on the calculated MSE. Repeat the above training process for multiple rounds of iterative training, and regularly use the validation set to evaluate the model performance. If the MSE of the validation set stops decreasing or starts to increase for several consecutive rounds, or other performance indicators show that the model is no longer optimized, terminate the training in advance. After the training is completed, the trained model is used as a reactive power prediction unit. Repeat the training L times (different initialization parameters or data subsets can be used) to obtain L reactive power prediction units, which are combined in parallel to form a reactive power prediction model. When the characteristic data of the power supply end is input, the L prediction units are calculated in parallel, and the comprehensive prediction results are used to obtain the real-time reactive power data of the wind farm to provide support for subsequent regulation.

[0047] L reactive power prediction units are connected in parallel to successfully build a reactive power prediction model. This parallel structure fully integrates the advantages of each prediction unit and enhances the model's adaptability and prediction ability to different wind farm partition characteristics.

[0048] The power supply characteristic data is fully input into the constructed reactive power prediction model. The model makes a deep prediction of the reactive power of each wind farm partition based on the pre-trained parameters and algorithm logic. After a complex calculation process, the real-time reactive power data is finally output, including the reactive power data of the L group partitions, which accurately reflects the real-time reactive power status of each wind farm partition in the current and future period, providing key data support for the dynamic regulation of wind farm reactive power, and helping to ensure the stable operation of wind farms and power grids.

[0049] In a possible implementation, step S500 further includes:

[0050] Step S510: Decompose the reactive power data of the L group partitions to obtain the upper limits of the P single-machine outputs of the P wind turbines. Step S520: Interactively obtain the lower limits of the P single-machine outputs of the P wind turbines. Step S530: Take the upper limits and lower limits of the P single-machine outputs as reactive power regulation constraints, update the upper limits of the P single-machine outputs, and obtain P updated single-machine output strategies. Step S540: Generate a first SVG regulation strategy based on the real-time reactive power demand and the P updated single-machine output strategies. Step S550: Use the P updated single-machine output strategies and the first SVG regulation strategy to simulate reactive power regulation, and obtain the SVG operation loss coefficient, grid voltage stability, and L partition stability coefficients of the L wind farm partitions. Step S560: Weightedly evaluate the SVG operation loss coefficient, grid voltage stability, and L partition stability coefficients, and output a first regulation stability. Step S570: Preset a regulation stability threshold. Step S580: Adjust the P updated single-machine output strategies and the first SVG regulation strategy according to the deviation percentage between the regulation stability threshold and the first regulation stability, and perform regulation stability evaluation based on the adjustment result until a reactive power distribution strategy is obtained in which the regulation stability data satisfies the regulation stability threshold, wherein the reactive power distribution strategy includes P single-machine power output strategies and an SVG power regulation strategy.

[0051] Specifically, the acquired reactive power data of L groups of partitions are deeply analyzed and processed to accurately determine the upper limit of the single-unit output of P wind turbines. First, according to the electrical topology of the wind farm, the connection relationship between each wind turbine and the partition is clarified, and the specific wind turbines contained in each partition are determined. Then, the power tracking decomposition algorithm is applied to each group of partition reactive power data. This algorithm is based on circuit principles and power transmission theory. By tracking the flow and distribution of reactive power in the power grid, it comprehensively considers multiple factors such as the location of wind turbines in the power grid, line impedance, power transmission characteristics, and mutual influence between wind turbines. Through a complex calculation process, the partition reactive power data is decomposed according to the contribution ratio of each wind turbine in the partition. For each partition, the maximum reactive power value that each wind turbine can theoretically output under the current partition reactive power distribution is calculated. These values ​​are the upper limit of the single-unit output of P wind turbines. The determination of this upper limit provides a key boundary reference for the subsequent reactive power control of wind turbines, ensuring that the control strategy is formulated within the operating capacity of the wind turbines, ensuring the safe and stable operation of the wind turbines, and also laying the foundation for the optimal allocation of reactive power in the entire wind farm.

[0052] By interactively communicating with the monitoring system and equipment management system of the wind turbine, the output lower limits of P individual wind turbines are obtained. This lower limit specifies the minimum reactive power output required for the stable operation of the wind turbine, ensuring that the wind turbine will not affect its normal operation and the stability of the entire wind farm during the regulation process due to too low reactive power output.

[0053] The P single-machine output upper limits and P single-machine output lower limits are used as the constraints for reactive power regulation, and the P single-machine output upper limits are optimized and updated. The optimization algorithm is used to comprehensively consider the operating efficiency of the wind turbine, equipment life, current wind speed and direction and other environmental factors as well as the real-time demand of the power grid for reactive power. Under the premise of meeting the constraints, P updated single-machine output strategies are calculated. These strategies are designed to enable wind turbines to provide reasonable reactive power support to the power grid as much as possible on the basis of safe and stable operation, and improve the overall operating performance of the wind farm.

[0054] According to the real-time reactive power demand and P updated single-machine output strategies, the reactive power balance principle is used to generate the first SVG regulation strategy. First, the real-time reactive power demand is carefully analyzed to clarify the precise reactive power demand of the current power grid as a whole and the changing trend of the demand. At the same time, the P updated single-machine output strategies are studied in depth to master the reactive power output plan of each wind turbine after optimization and adjustment. Then, based on the principle that reactive power must be balanced in the power grid, the difference between the current reactive power output of the wind turbine and the reactive power demand of the power grid is calculated. Considering that the SVG device has fast and flexible reactive power regulation capabilities, this difference is used as the reactive power target value that the SVG device needs to compensate or absorb. Then, based on the performance parameters, adjustment range, response speed and other characteristics of the SVG device itself, combined with the operating constraints of the power grid, such as voltage upper and lower limits, harmonic content limits, etc., the optimization algorithm is used to formulate the first SVG regulation strategy. The strategy specifies in detail how the SVG device should adjust its reactive power output size, phase, and regulation rate under current operating conditions to ensure that the SVG device can work in coordination with the wind turbine, accurately meet the grid's real-time demand for reactive power, maintain grid voltage stability, and improve the overall effect of wind farm reactive power regulation.

[0055] With the help of a high-precision wind farm power system simulation model, the reactive power regulation when executing P updated single-machine output strategies and the first SVG regulation strategy is simulated to obtain the key performance coefficient. First, the P updated single-machine output strategies and the first SVG regulation strategy are accurately input into the simulation model. The model simulates the dynamic process of the wind turbine operating according to the updated single-machine output strategy based on the detailed electrical parameters, connection relationships and operation characteristics of the power grid of each device in the wind farm, and the SVG device performs reactive power compensation or absorption according to the first SVG regulation strategy. In the simulation process, the SVG operation loss coefficient is calculated through the detailed modeling and analysis of the power conversion process inside the SVG device. This coefficient accurately reflects the energy loss of the SVG device under the current regulation strategy, and its size depends on factors such as the topological structure of the SVG, the characteristics of the components and the frequency of reactive power regulation. At the same time, the voltage changes of each node in the power grid are closely monitored, and the voltage stability of the power grid is calculated using the voltage stability analysis algorithm. This indicator comprehensively evaluates the influence of the regulation strategy on the voltage stability of the power grid by considering factors such as the amplitude and frequency of voltage fluctuations and the correlation between voltages at different nodes. In addition, for the L wind farm partitions, the partition stability evaluation model is used to calculate the L partition stability coefficients based on the actual operating status of the wind turbines in each partition, the reactive power output, and the power interaction relationship with other partitions. This coefficient comprehensively considers multiple factors such as power balance within the partition, equipment overload risk, and reactive power reserve, aiming to ensure that each partition can maintain stable operation during the reactive power regulation process, and avoid the situation where some partitions are overloaded with reactive power, causing equipment overheating and damage, or underload affecting the overall performance of the power grid. Through this series of simulation calculations, scientific and accurate data support is provided for the comprehensive evaluation of the performance of the regulation strategy.

[0056] The simulated SVG operation loss coefficient, grid voltage stability and L partition stability coefficients are weighted and evaluated. According to the operation objectives and actual needs of the wind farm, the corresponding weights are set for each coefficient, and the first regulation stability is obtained by comprehensive calculation. This indicator fully reflects the overall performance of the current reactive power regulation strategy, covering multiple aspects such as economy and stability.

[0057] According to the design standards, operation experience and grid requirements of the wind farm, a control stability threshold is preset. This threshold is the key criterion for judging whether the control strategy meets the requirements, and clarifies the minimum level that the wind farm reactive power control should reach in terms of stability and economy.

[0058] The deviation percentage between the control stability threshold and the first control stability is accurately calculated. This deviation percentage directly reflects the degree of difference between the current control strategy and the ideal state. If the deviation percentage exceeds the acceptable range, it indicates that the current strategy needs to be optimized. Then, based on the deviation information, the intelligent feedback adjustment algorithm is used to fine-tune the P updated single-machine output strategies and the first SVG adjustment strategy. For the single-machine output strategy, it involves fine-tuning the upper and lower limits of the reactive power output of the wind turbine, comprehensively considering the operating efficiency, heating conditions and contribution to the stability of the power grid of the unit, and redistributing the reactive power tasks of each unit. For the SVG adjustment strategy, according to the real-time operating status of the SVG, the remaining adjustment capacity and the voltage fluctuation of the power grid, the size, phase and response rate of its reactive power compensation are optimized. After the adjustment is completed, the new strategy combination is substituted into the high-precision wind farm power system simulation model again to conduct a new round of control stability evaluation. Repeat the above steps of calculating deviation, adjusting strategy and evaluating stability, and continue to iterate. Each round of iteration moves in the direction of reducing deviation and approaching the control stability threshold until the reactive power allocation strategy whose control stability data meets the preset control stability threshold is finally obtained. The P single-machine power output strategies and SVG power regulation strategies determined at this time can ensure that the wind farm can achieve efficient, stable and economical operation while meeting the reactive power demand of the power grid, laying a solid foundation for the sustainable development of the wind farm.

[0059] In a possible implementation, step S340 further includes:

[0060] Step S341: Call the local historical power consumption data of the multiple distributed power supply areas to obtain the multiple partition power consumption characteristics. Step S342: Perform permutation and combination enumeration on the multiple distributed power supply areas, obtain K power supply area pairs, and then calculate the Euclidean distance of the multiple partition power consumption characteristics according to the K power supply area pairs to obtain K regional similarity coefficients. Step S343: Take the multiple distributed power supply areas as topological nodes, take the K regional similarity coefficients as the topological connection length, connect the multiple distributed power supply areas according to the K power supply area pairs, and obtain regional similarity feature topology. Step S344: Preset a regional similarity threshold, and split the regional similarity feature topology into M local similarity feature topologies by deleting the topological connection of the regional similarity feature topology according to the regional similarity threshold. Step S345: Extract topological nodes from the M local similarity feature topologies to obtain the M groups of distributed power supply areas.

[0061] Specifically, in the initial stage of planning and analyzing the power supply area of ​​a wind farm, it is crucial to accurately grasp the power consumption characteristics of each area. First, based on the established power grid structure of the wind farm, the power supply range covered by different lines and substations is clarified, and on this basis, the wind farm is divided into multiple distributed power supply areas. Subsequently, in-depth calls for local historical power consumption data are carried out for these areas. With the help of advanced data management systems, detailed records related to power consumption in each region are accurately screened out from massive storage data, covering power consumption information in different time periods, seasons and various working conditions. Further subdivided according to load characteristics, considering the proportion of residential power consumption, industrial power consumption, commercial power consumption, etc. in each region, analyzing the time pattern of load peaks, and the fluctuation of loads with external factors (such as weather changes, holidays, etc.). Using data processing algorithms, these data are sorted and integrated, and finally multiple partition power consumption characteristics are accurately extracted. These characteristics will lay a solid data foundation for subsequent work such as building regional similar feature topologies and formulating reactive power control strategies, ensuring that wind farm operation and management can closely fit the actual power demand of each region and achieve efficient and stable power supply.

[0062] A comprehensive and rigorous permutation and combination enumeration operation is implemented for many distributed power supply areas to obtain K power supply area pairs. Based on the obtained pairs, in order to accurately quantify the similarity of the power consumption characteristics between them, the Euclidean distance calculation model is used based on the power consumption characteristics of multiple partitions previously obtained. With the help of mathematical methods, the power consumption characteristic vectors of each pair of power supply areas are accurately evaluated in multi-dimensional space to obtain K regional similarity coefficients. These coefficients intuitively present the degree of similarity or difference in power consumption characteristics between different regional pairs in the form of quantitative values, providing key quantitative support for the subsequent construction of regional association architecture.

[0063] The topology construction method is adopted to cleverly transform multiple distributed power supply areas into topological nodes. These nodes are like key hubs in the network, each carrying the core power consumption characteristic information of the corresponding area, such as the unique load change trend, power consumption type composition, etc., laying a solid foundation for the subsequent construction of topological relationships. At the same time, the K regional similarity coefficients calculated rigorously before are used as the length of the topological connection line. From the perspective of the integration of mathematical and physical meanings, the smaller the regional similarity coefficient, the closer the power consumption characteristics of the corresponding areas are, which is reflected in the topological map as the shorter the connection line connecting them; conversely, the longer the connection line, the closer the degree of connection between the regions is. Then, strictly in accordance with the logical links set by the pairing of K power supply areas, professional topological drawing algorithms or tools are enabled to accurately build topological connections between the distributed power supply areas. The two ends of a connection line are connected to two areas with a specific similarity coefficient. The final regional similarity characteristic topological architecture provides a very critical visualization basis and analysis basis for the subsequent work such as reactive power regulation based on regional clustering, and strongly promotes the refinement of wind farm operation and management.

[0064] Based on the in-depth mining of past operation data of wind farms, simulation analysis of a large number of actual working conditions, comprehensive consideration of wind farm power consumption characteristics, regulation accuracy requirements, equipment operation stability and other factors, the regional similarity threshold is determined, and the constructed regional similarity feature topology is finely processed. For each line in the topology, the corresponding regional similarity coefficient must be compared with the preset threshold one by one. If the similarity coefficient of the area connected by a line is less than or equal to the threshold, it means that the two areas have sufficient similarity in power consumption characteristics and meet the conditions for further clustering and integration, and the line is retained; on the contrary, if the similarity coefficient is greater than the threshold, it means that the power consumption of the two areas is quite different. In order to avoid the increase in the complexity of subsequent analysis and regulation, this line will be deleted. Through such orderly screening and deletion operations, the originally complete but possibly redundantly connected regional similarity feature topology is cleverly split into M local similarity feature topologies. The regions within each local similar feature topology have highly similar electricity consumption characteristics, forming a tight clustering group. This not only reduces the difficulty of subsequent analysis, but also provides a clear framework and solid foundation for more accurate formulation of reactive power control strategies for different regional clusters and optimization of resource allocation, making wind farm operation and management more targeted and gradually moving towards efficiency and intelligence.

[0065] After successfully splitting the regional similar feature topology into M local similar feature topologies, the next step is to focus on these local structures. Since each local similar feature topology is composed of a series of topological nodes with specific connection relationships, and these nodes essentially correspond to the actual distributed power supply area. At this time, a special topological node extraction algorithm is used to accurately identify and separate the nodes in each local similar feature topology. These extracted nodes are recombined to form M groups of distributed power supply areas based on their original corresponding relationships. In this way, the M groups of distributed power supply areas obtained show extremely high power consumption similarity between the internal areas, which provides a basis for accurate classification for subsequent work such as building a highly targeted reactive power prediction model and implementing a control strategy that fits the regional characteristics, so that the wind farm can achieve differentiated and refined operations based on regional characteristics during the operation and management process, thereby ensuring the stability and efficiency of the power supply of the entire wind farm, and effectively promoting the wind farm to stride forward towards intelligent operation and maintenance.

[0066] In a possible implementation, step S341 further includes:

[0067] The regional power consumption characteristics include the total regional load, the regional load type ratio, the regional load standard deviation, the regional load change rate and the regional load fluctuation ratio.

[0068] Specifically, the extracted regional power consumption characteristics contain multiple core dimensions, providing a key basis for precise control. The total regional load directly reflects the total reactive power demand of all users in the region, and presents the reactive power scale that the region needs to bear during peak power consumption periods with specific values. It is an important indicator for measuring power supply pressure. The regional load type ratio accurately divides the load proportions of industrial, commercial, and residential users, thereby clarifying the dominant type of regional power consumption. For example, in areas with a high industrial proportion, its power consumption period and fluctuation law are directly related to the industrial production process; in areas where commerce is dominant, the load peak is concentrated during business hours. The regional load standard deviation calculates the degree to which the load data deviates from the average value. The larger the standard deviation value, the lower the power stability in the region and the more obvious the load fluctuation. The regional load change rate reflects the changing trend of power demand in real time based on the increase or decrease value of the load in unit time, and assists in predicting the direction of power supply pressure changes. The regional load fluctuation ratio is reflected by the ratio of peak load demand to valley load demand. The larger the ratio, the more prominent the difference between peak and valley power consumption, and the higher the requirements for power supply regulation capability. Combining these characteristics, we can gain a comprehensive insight into the electricity consumption of each zone, laying a solid data foundation for subsequent work such as building regional similar feature topologies and formulating reactive power control strategies, ensuring that wind farm operation and management are in line with the actual electricity demand of each region and achieving efficient and stable power supply.

[0069] In a possible implementation, step S350 further includes:

[0070] Step S351: Call to obtain the first group of historical reactive power demands and the first historical external meteorological characteristics of the first group of distributed power supply areas. Step S352: Decompose the first group of historical reactive power demands using the demand-side sliding window to obtain H groups of stage reactive power demands, perform intra-group time series mapping and summation on the H groups of stage reactive power demands, and obtain H stage reactive power sequences. Step S353: Decompose the first historical external meteorological characteristics using the demand-side sliding window to obtain H stage external meteorological sequences, wherein the H stage external meteorological sequences have H stage time feature coding identifiers. Step S354: Collect and obtain H time window closed points of the H stage reactive power sequences. Step S355: Take 1 / t of the demand-side sliding window as the prediction time window. Step S356: Take the H time window closed points as the data extraction starting point, perform data segmentation on the H stage reactive power sequences according to the prediction time window, and obtain H updated reactive power sequences. Step S357: Use the H-stage external meteorological sequences, H-stage time feature codes and H-stage reactive power sequences as training data to construct a first dynamic reactive power prediction model. Step S358: Similarly, construct M dynamic reactive power prediction models for the M groups of distributed power supply areas, and complete the construction of the reactive power prediction collaborative model by connecting the M dynamic reactive power prediction models in parallel.

[0071] Specifically, given that the first group of distributed power supply areas are geographically close and are consistently affected by external meteorological factors, the first group of historical reactive power demands and the first historical external meteorological characteristics corresponding to the area in the database are accurately called up. These data record the reactive power consumption of regional electricity in different periods in the past and the meteorological conditions at the time, providing basic material for subsequent model construction.

[0072] Start the data processing process. Use the demand-side sliding window technology to decompose the first group of historical reactive power demands, and accurately divide H groups of reactive power demands according to specific time segments. Then, perform the intra-group time series mapping and summing operation on these H groups of data. According to the inherent logic of the time series, reasonably superimpose the reactive power values ​​at different times in each group, and finally successfully obtain H phase reactive power sequences. These sequences show the changing pattern of reactive power demand from a phased perspective.

[0073] A similar operation is performed on the first historical external meteorological feature. Similarly, with the help of the demand-side sliding window, it is disassembled into H stages of external meteorological sequences, and each sequence is assigned H stage time feature coding identifiers, so that the meteorological information and reactive power demand in different periods can be accurately matched in the future, and the potential relationship between the two can be explored.

[0074] Through the data acquisition algorithm, H time window closed points of H phase reactive power sequences are accurately locked and acquired from massive data. These closed points, as key time nodes, play an important positioning role in subsequent data segmentation and model construction.

[0075] Clarify the forecast time window setting rules and set 1 / t of the demand-side sliding window as the forecast time window. This means that the model inputs data from the past t time to accurately predict data from the future 1 / t time. By setting the time window reasonably, the model's reliance on historical data and its ability to predict future trends can be balanced.

[0076] Taking the H time window closure points as the starting point for data extraction, the H stage reactive power sequences are finely segmented according to the set prediction time window. According to the established rules, the subsequences used for model training are accurately selected, and finally H updated reactive power sequences are obtained to ensure that the data input into the model is highly targeted and timely.

[0077] The H-stage external meteorological sequences, H-stage time feature codes, and H-stage reactive power sequences obtained through careful processing in the early stage are integrated as complete training data and put into the construction process of the first dynamic reactive power prediction model. The convolutional neural network and long short-term memory network algorithms are used to train the model based on these data, so that it has the ability to accurately predict future reactive power based on meteorological characteristics and historical reactive power changes.

[0078] Similarly, M dynamic reactive power prediction models corresponding to M groups of distributed power supply areas are constructed one by one. In addition, in order to give full play to the advantages of multi-regional data, these M dynamic reactive power prediction models are connected in parallel to achieve data sharing and collaborative computing, and finally complete the construction of the reactive power prediction collaborative model. This model can comprehensively consider the characteristics of different regions, accurately predict the reactive power demand of the entire wind farm, provide solid theoretical support and data guarantee for subsequent reactive power regulation, and ensure the stable and efficient operation of the wind farm.

[0079] In a possible implementation, step S357 further includes:

[0080] Step S3571: Local feature identification is performed on the H-stage external meteorological sequences, the H-stage time feature codes and the H-stage reactive power sequences to obtain the H-stage output data features. Step S3572: The H-stage external meteorological sequences and the H-stage reactive power sequences are used as input data, and the H-stage output data features are used as output results to perform model training of the time series feature extraction unit, wherein the time series feature extraction unit is constructed based on the CNN model. Step S3573: The H-stage output data features are used as input data, and the H-stage updated reactive power sequences are used as output results to perform model training of the time series power prediction unit, wherein the time series power prediction unit is constructed based on the LSTM model. Step S3574: The output end of the time series feature extraction unit is connected to the input end of the time series power prediction unit, and after the construction of the first dynamic reactive power prediction model is completed, the H-stage external meteorological sequences, the H-stage time feature codes, the H-stage reactive power sequences and the H-stage output data features are used to perform joint training of the first dynamic reactive power prediction model.

[0081] Specifically, in the local feature identification stage, a data identification algorithm is adopted. This algorithm accurately identifies the key local features in the H-stage external meteorological sequences, H-stage time feature codes and H-stage reactive power sequences based on the physical meaning of the data, the timing characteristics and the importance of the subsequent model training. The key turning points of the reactive power sequence are captured according to the characteristics of the power system, so as to extract the output data features of the H stages.

[0082] The time series feature extraction unit is built based on the CNN (convolutional neural network) model. The CNN algorithm takes advantage of its core convolutional layer and pooling layer architecture. The convolutional layer slides on the input data (H stages of external meteorological sequences and H stages of reactive power sequences) through convolution kernels of different sizes to automatically extract local features in the data. The pooling layer further reduces the dimension and abstracts these features, thereby efficiently capturing complex time series features, such as the gradual trend of meteorology and the periodic fluctuation of reactive power. The H-stage output data features are used as the output results for training, so that the model has accurate time series feature perception capabilities.

[0083] The time series power prediction unit is constructed based on the LSTM (Long Short-Term Memory Network) model. The LSTM algorithm operates with a unique memory unit and gating mechanism. The memory unit stores the key information of the past input sequence, and the gating mechanism controls the retention and update of the information. When the H-stage output data features are input, it can fully consider the correlation between the long sequence data and the past and future, and accurately calculate the reactive power trend in different periods in the future based on these condensed key information. The model is trained with H updated reactive power sequences as the output results, providing a reliable forward-looking basis for reactive power regulation.

[0084] The output end of the time series feature extraction unit built based on the CNN model is precisely connected to the input end of the time series power prediction unit built based on the LSTM model. Through such a connection, the complex time series features extracted from the input data can be smoothly transmitted to the module responsible for predicting the future reactive power trend, completing the preliminary architecture of the first dynamic reactive power prediction model, so that the model has the initial ability to integrate multiple information for prediction. Subsequently, the crucial joint training process was started, making full use of the H-stage external meteorological sequences carefully prepared in the early stage, which cover key meteorological elements such as temperature, air pressure, wind speed in different periods, to provide the model with information on changes in the external environment; H-stage time feature encoding, so that the model can clearly identify the time series position of each data; H-stage reactive power sequences, recording the actual change trajectory of reactive power in the past; and H-stage output data features, which are the core feature sets that have been finely identified and refined in the early stage. These four sets of data are input into the preliminarily constructed model in an all-round and repeated manner. The model predicts based on the input data and calculates the prediction error by comparing it with the actual historical data. Based on error feedback, the optimization algorithm is used to fine-tune the model's weights, biases and other parameters. After multiple rounds of such input, prediction, feedback and adjustment processes, the internal structure and parameter settings of the model are continuously optimized to ensure that the model can deeply integrate multi-dimensional data information, accurately adapt to the complex and changeable reactive power prediction needs of wind farms, and effectively improve the accuracy and reliability of reactive power prediction, providing solid technical support for the efficient and stable operation of subsequent wind farms.

[0085] It should be noted that the above-mentioned sequence of the embodiments of the present application is only for description and does not represent the advantages and disadvantages of the embodiments. And the above-mentioned specific embodiments of this specification are described. In addition, the processes depicted in the accompanying drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0086] The above description is only a preferred embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present application should be included in the protection scope of the present application.

[0087] This specification and the drawings are merely exemplary illustrations of the present application and are deemed to cover any and all modifications, variations, combinations or equivalents within the scope of the present application. Obviously, a person skilled in the art may make various modifications and variations to the present application without departing from the scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the present application and its equivalents, the present application intends to include these modifications and variations.

Claims

1. A wind farm reactive power control method based on source-load characteristic data analysis, characterized in that: The method comprises: Preset demand-side sliding window and supply-side sliding window; The demand-side sliding window and the power supply-side sliding window are used as data collection time limits to collect source-load characteristic data and obtain demand-side characteristic data and power supply-side characteristic data; Perform regional distributed reactive power demand forecasting on the demand-side characteristic data to obtain real-time reactive power demand; Based on the power supply end characteristic data, reactive power prediction of wind farm partitions is performed to obtain real-time reactive power data; The reactive power flow cloud receives and optimizes reactive power partition control according to the real-time reactive power demand and the real-time reactive power data to obtain a reactive power allocation strategy; The reactive power distribution strategy is loaded into the reactive power regulation device and the SVG device of the target wind farm, and the reactive power of the wind farm is dynamically regulated through the reactive power regulation device and the SVG device in coordination.

2. The method for controlling reactive power of a wind farm by analyzing source-load characteristic data according to claim 1, characterized in that: Performing regional distributed reactive power demand forecasting on the demand-side characteristic data to obtain real-time reactive power demand, the method comprising: interactively obtaining power supply coverage information of the target wind farm, and screening and generating a target power grid topology according to the power supply coverage information; After substation positioning is performed in the target power grid topology, the target power grid topology is divided into a plurality of distributed power supply areas according to current flow directions; Locally call to obtain multiple partition power consumption characteristics of the multiple distributed power supply areas; Performing regional cluster analysis according to the power consumption characteristics of multiple partitions of the multiple distributed power supply areas to obtain M groups of distributed power supply areas; Constructing M dynamic reactive power prediction models for the M groups of distributed power supply areas, and completing the construction of a reactive power prediction collaborative model by connecting the M dynamic reactive power prediction models in parallel; The reactive power prediction collaborative model uses the demand-side characteristic data as input to perform regional distributed reactive power demand prediction and outputs the real-time reactive power demand.

3. The wind farm reactive power control method based on source-load characteristic data analysis according to claim 1, characterized in that: Based on the power supply end characteristic data, reactive power prediction of wind farm partitions is performed to obtain real-time reactive power data, and the method includes: Interactively obtaining power generation characteristics of P wind turbines in the target wind farm; Performing a topological analysis on the P wind turbines based on the power generation characteristics of the P wind turbines, and dividing the P wind turbines into L wind farm partitions; Calling historical data of the L wind farm partitions to obtain L groups of historical power generation data; Using 1 / y of the power supply end sliding window as a data segmentation scale, decomposing the L groups of historical power generation data, and obtaining L groups of historical reactive power sequences; Collect wind power environment data according to the L group terminal time nodes of the L group historical reactive power sequences and the power supply end sliding window to obtain L group wind power environment data sequences; A standard prediction model is pre-built using an LSTM model, wherein the input data of the standard prediction model is a wind power environment data sequence, the output data of the standard prediction model is a reactive power data sequence, and the time ratio of the output result to the input result is 1 / y; Using the L groups of wind power environment data sequences and the L groups of historical reactive power sequences as training data to perform directional parameter update of the standard prediction model, and obtaining L reactive power prediction units, wherein the L reactive power prediction units include L groups of reactive power prediction branches connected in parallel; The L reactive power prediction units are connected in parallel to complete the construction of the reactive power prediction model; The power supply end characteristic data is input into the reactive power prediction model to perform wind farm partition reactive power prediction to obtain the real-time reactive power data, wherein the real-time reactive power data includes L-group partition reactive power data.

4. The method for controlling reactive power of a wind farm by analyzing source-load characteristic data according to claim 3, characterized in that: The reactive power flow cloud receives and optimizes reactive power partition control according to the real-time reactive power demand and the real-time reactive power data to obtain a reactive power allocation strategy. The method includes: Decomposing the L group partition reactive power data to obtain P single-unit output upper limits of the P wind turbines; Interactively obtaining P single-machine output lower limits of the P wind turbines; Taking the P single-machine output upper limits and the P single-machine output lower limits as reactive power regulation constraints, updating the P single-machine output upper limits to obtain P updated single-machine output strategies; Generate a first SVG regulation strategy according to the real-time reactive power demand and P updated single-machine output strategies; Using the P updated single-machine output strategies and the first SVG regulation strategy to perform reactive power control simulation, to obtain the SVG operation loss coefficient, the grid voltage stability and the L partition stability coefficients of the L wind farm partitions; Weighted evaluation of the SVG operation loss coefficient, grid voltage stability and L partition stability coefficients, and outputting a first regulation stability; Preset control stability threshold; The P updated single-machine output strategies and the first SVG regulation strategy are adjusted according to the deviation percentage between the regulation stability threshold and the first regulation stability, and the regulation stability is evaluated based on the adjustment result until a reactive power distribution strategy is obtained in which the regulation stability data satisfies the regulation stability threshold, wherein the reactive power distribution strategy includes P single-machine power output strategies and an SVG power regulation strategy.

5. The method for controlling reactive power of a wind farm by analyzing source-load characteristic data according to claim 2, characterized in that: Performing regional cluster analysis according to the power consumption characteristics of multiple partitions of the multiple distributed power supply areas to obtain M groups of distributed power supply areas, the method comprising: Calling local historical power consumption data of the multiple distributed power supply areas to obtain the power consumption characteristics of the multiple partitions; After enumerating the multiple distributed power supply areas by permutations and combinations to obtain K power supply area pairs, Euclidean distance calculation of the multiple partition power consumption characteristics is performed according to the K power supply area pairs to obtain K area similarity coefficients; Taking the multiple distributed power supply areas as topological nodes, taking the K regional similarity coefficients as topological connection lengths, and connecting the multiple distributed power supply areas according to the pairing of the K power supply areas to obtain a regional similarity feature topology; Preset a regional similarity threshold, and perform topological line deletion of the regional similar feature topology according to the regional similarity threshold, so as to split the regional similar feature topology into M local similar feature topologies; Topological nodes are extracted from the M local similar characteristic topologies to obtain the M groups of distributed power supply areas.

6. The method for controlling reactive power of a wind farm by analyzing source-load characteristic data according to claim 5, characterized in that: The regional power consumption characteristics include the total regional load, the regional load type ratio, the regional load standard deviation, the regional load change rate and the regional load fluctuation ratio.

7. The method for controlling reactive power of a wind farm by analyzing source-load characteristic data according to claim 2, characterized in that: Constructing M dynamic reactive power prediction models for the M groups of distributed power supply areas, and completing the construction of a reactive power prediction collaborative model by connecting the M dynamic reactive power prediction models in parallel, the method comprising: Calling to obtain a first group of historical reactive power demands and a first group of historical external meteorological characteristics in a first group of distributed power supply areas; Decomposing the first group of historical reactive power demands by using the demand-side sliding window to obtain H groups of stage reactive power demands, performing intra-group time series mapping and summing on the H groups of stage reactive power demands to obtain H stage reactive power sequences; Decomposing the first historical external meteorological feature by using the demand-side sliding window to obtain H stage external meteorological sequences, wherein the H stage external meteorological sequences have H stage time feature coding identifiers; Collecting and obtaining H time window closing points of the H stage reactive power sequences; Taking 1 / t of the demand-side sliding window as the prediction time window; Taking the H time window closed points as data extraction starting points, performing data segmentation on the H stage reactive power sequences according to the prediction time window, and obtaining H updated reactive power sequences; The H-stage external meteorological sequences, the H-stage time feature codes and the H-stage reactive power sequences are used as training data to construct a first dynamic reactive power prediction model; By analogy, M dynamic reactive power prediction models of the M groups of distributed power supply areas are constructed, and the construction of the reactive power prediction collaborative model is completed by connecting the M dynamic reactive power prediction models in parallel.

8. The method for controlling reactive power of a wind farm by analyzing source-load characteristic data according to claim 7, characterized in that: The H-stage external meteorological sequences, the H-stage time feature codes and the H-stage reactive power sequences are used as training data to construct a first dynamic reactive power prediction model, and the method includes: Performing local feature identification on the H-stage external meteorological sequences, the H-stage time feature codes, and the H-stage reactive power sequences to obtain H-stage output data features; Taking the H-stage external meteorological sequences and the H-stage reactive power sequences as input data, and taking the H-stage output data features as output results, a model training of a time series feature extraction unit is performed, wherein the time series feature extraction unit is constructed based on a CNN model; Taking the H stage output data features as input data and the H updated reactive power sequences as output results, model training of a time series power prediction unit is performed, wherein the time series power prediction unit is constructed based on an LSTM model; The output end of the time series feature extraction unit is connected to the input end of the time series power prediction unit. After completing the construction of the first dynamic reactive power prediction model, the H stage external meteorological sequences, H stage time feature codes, H stage reactive power sequences and H stage output data features are used to perform joint training of the first dynamic reactive power prediction model.

Images