CN115390448A - Visual analysis method and system for control strategy of coal-fired power plant - Google Patents

Abstract

The invention discloses a visual analysis method for a control strategy of a coal-fired power plant, which comprises the following steps: step 1, obtaining historical operation data of a coal-fired power plant, wherein the historical operation data comprises equipment data acquired by a sensor and corresponding sensor position data; step 2, analyzing the equipment data in a preset time interval on the basis of the sensor position data to acquire control strategy information in the time interval; and 3, importing the equipment data, the sensor position data and the control strategy information into a pre-constructed visual model, carrying out graphic drawing on the imported data and information in the visual model, carrying out interactive operation on the data in the graphic, and then outputting and displaying the data. The invention also provides a visual analysis system based on the method. The method allows a user to specify a query control strategy, provides an algorithm for automatically matching and mining the control strategy, and supports the spatial propagation mode and influence time lag analysis of the control strategy.

Description

A visual analysis method and system for control strategies of coal-fired power plants

technical field

The invention belongs to the technical field of control strategy analysis, and in particular relates to a visual analysis and system for the control strategy of a coal-fired power plant.

Background technique

Control strategy is an important factor affecting the efficiency of coal-fired power generation. Experts often follow specific control strategies when regulating coal-fired power plants to achieve corresponding goals, such as increasing the combustion rate and reducing pollution emissions. A large number of sensors monitor the operation of coal-fired power plants. These sensors can be divided into two categories according to the monitoring categories: control sensors that monitor control variables (such as valves), and state sensors that monitor physical states (such as temperature). The changes detected by these sensors are defined as corresponding control events and status events. A control strategy is a sequence of related control events and status events.

Control strategy analysis techniques fall into two categories: experience-driven methods and data-driven methods.

The analytical process supported by the experience-driven approach includes data collection, condition monitoring, and data presentation. For example, RSSFAD collects the data continuously transmitted by the sensor and visualizes it through a simple dashboard. The characteristics of this method are: first, it has good real-time performance and can support online monitoring of electric field data; second, it shows fine granularity , using visualization to present the original time series data; third, the automatic algorithm without inheritance does not support the extraction of control strategies, and needs to analyze the original data based on experience to judge the control strategy, so it is impossible to complete the mining and analysis of control strategies for large-scale data. Moreover, the presentation methods of the analysis results are mostly simple static charts, for example, RSSFAD draws a simple process structure. These single presentation methods limit users' in-depth understanding of coal-fired power plant control strategies.

Existing data-driven methods first model coal-fired power plants and then simulate control strategies, supporting modeling of individual units or entire units. When modeling individual units, the goal is to optimize the control logic within a single unit, and the complete control strategy across units cannot be analyzed. When modeling the overall unit, strict modeling conditions (such as steady-state operation) must be set, otherwise the derivation cannot be established, or machine learning based on historical data requires huge data accumulation and reasonable interpretable rules. These conditions limit the scalability of the corresponding methods.

Patent document CN112102111A discloses a power plant data intelligent processing system, including: including: operation and maintenance management module; safety management module; data acquisition and extraction module; data storage module; general analysis and calculation module; intelligent calculation and analysis module: based on Machine learning algorithm, using distributed computing resources to establish and train intelligent algorithm models, and building prediction models through multi-dimensional associated data analysis data; edge computing module; microservice publishing module: used to publish and manage system load data, computing tasks and external Services; Data Governance Module. But this method ignores the time-lag problem existing in sensor feedback.

Contents of the invention

In order to solve the above problems, the present invention provides a visual analysis method for the control strategy of the coal-fired power plant. Based on the influence of the control strategy on the coal-fired power plant, the abnormal events of the coal-fired power plant are processed by using time-delay relationship optimization.

A visual analysis approach for coal-fired power plant control strategies, including:

Step 1. Obtain the historical operation data of the coal-fired power plant, the historical operation data including the equipment data collected by the sensor and the corresponding sensor location data;

Step 2. Based on the sensor position data, analyze the device data within a preset time interval, and obtain control strategy information within the time interval;

Step 3. Import device data, sensor location data and control strategy information into the pre-built visualization model, in the visualization model, draw graphics for the imported data and information, and perform interactive operations on the data in the graphics The output is displayed.

Specifically, the device data includes device operation status data and control command data.

Specifically, the time interval takes the range time of power generation efficiency, furnace negative pressure parameters, or sudden changes in pollutant discharge, that is, the area containing the maximum or minimum value of the first-order derivative, and the threshold is set to 80% of the maximum absolute value of the first-order derivative. %.

Preferably, the analysis in step 2 adopts an aggregation operation method to analyze the spatial position of the sensors, the degree of correlation between adjacent sensors, and the relationship between sensor data changes and time series.

Preferably, the relationship between the sensor data change and the time series is displayed in alignment with time delay.

The present invention also provides a visual analysis system with simple operation and fast feedback, based on the above-mentioned visual analysis method for the control strategy of coal-fired power plants, including:

The filter view module, according to the occurrence of abnormal events, filters and analyzes the historical operation data, and outputs the time interval where the abnormal events are located;

The detailed view module performs an aggregation operation according to the time interval output by the filter view module and the historical operation data, and outputs the control strategy information in the time interval;

The graph view module analyzes and feeds back the propagation mode of the control strategy in space according to the obtained control strategy information;

The strategy view module draws the time cascade corresponding to the control strategy according to the obtained control strategy information;

The algorithm module calculates and outputs specific control strategies, propagation information, and time cascade information based on the interaction information and operating data of the above four modules.

Preferably, the analysis and filtering of the filter view module is based on a delay-aware control strategy extraction method, and the control strategy extraction method includes forward query and backward query.

Preferably, the forward query adopts the longest common subsequence algorithm, divides the abnormal events in the time interval into multiple rising and falling events, and performs fuzzy matching according to preset conditions.

Specifically, the specific process of the forward query is as follows:

Step 1. Discretize all historical operating data into trend intervals, namely rising intervals, declining intervals, and stable intervals;

Step 2. Segment and align all intervals, so that the historical operating data can be expressed as a matrix M _S , where each row represents a sensor, and each column represents a time interval. The values in the matrix indicate the specific time interval of the sensor in this time interval. Trend, 0 means stable, 1 means down, 2 means up;

Step 3. Record the control strategy of the forward query as a similar matrix M _T , that is, each row represents the sensors involved in the target control strategy, and each column represents the expected trend change;

Step 4. Take out the row where the sensor is shared with the matrix M _T from the matrix M _S to obtain M _A , make the number of rows of M _A and M _T the same, and then merge each column of the two matrices by row to obtain two sequences A and T;

Step 5, use the longest common subsequence matching algorithm for sequences A and T, F(i, j) represents the length of the longest common subsequence in the first i item of A and the first j item of T, and the dynamic programming transfer equation is as follows:

Preferably, the backward query performs extended retrieval based on key sensors that feed back abnormal events within a time interval.

Specifically, the specific process of the backward query is as follows:

Step 1. Add the key sensor to the search queue L, and set the time T to 0;

Step 2. Obtain the first sensor P in the search queue;

Step 3. Shift the time series of all sensors Q to be allocated backward by t time steps, and calculate the time series of sensors Q to be allocated and the first sensor P in the specified time zone for each t within 1 to 15 minutes Pearson coefficient c:

Step 4. If the Pearson coefficient c is greater than 0.8, it is determined to be relevant, and the sensor Q to be allocated is added to the search queue L, and t is added to the time T;

Step 5. If the search queue L is an empty set or the time T exceeds 2 hours, terminate the search; otherwise, go back to step 2 to continue.

Compared with prior art, the beneficial effect of the present invention:

Allows users to specify query control strategies, provides algorithms for automatic matching and mining of control strategies, supports spatial propagation modes and impact time-lag analysis of control strategies, does not limit the model and operating conditions of coal-fired power plants, and uses appropriate visualization from multi-level, It displays the analysis results of the control strategy from multiple angles, and allows users to interactively explore step by step and finally verify the credibility of the control strategy.

Description of drawings

Fig. 1 is a schematic diagram of a visual analysis method for a control strategy of a coal-fired power plant provided by the present invention;

FIG. 2 is a schematic diagram of a visual analysis system for a control strategy of a coal-fired power plant provided in this embodiment;

FIG. 3 is a schematic diagram of a filter view module provided in this embodiment;

FIG. 4 is a schematic diagram of a graph view module provided in this embodiment;

FIG. 5 is a schematic diagram of a policy view module provided in this embodiment;

FIG. 6 is a schematic diagram of a detail view module provided in this embodiment;

FIG. 7 is a schematic diagram of a forward query provided in this embodiment;

FIG. 8 is a flowchart of the backward query provided by this embodiment.

Detailed ways

As shown in Figure 1, a visual analysis method for the control strategy of a coal-fired power plant includes:

Step 1. Obtain the historical operation data of the coal-fired power plant, wherein the historical operation data includes the operation status data of the equipment, the control command data and the corresponding sensor position data;

Step 2. According to the range of abnormal conditions during the operation of the coal-fired power plant, the aggregation operation method is used to analyze the spatial position of the sensor, the correlation between adjacent sensors, and the relationship between sensor data changes and time series. Obtain control strategy information within the time interval;

Step 3. Import device data, sensor location data and control strategy information into the pre-built visualization model, in the visualization model, draw graphics for the imported data and information, and perform interactive operations on the data in the graphics The output is displayed.

The present invention also provides a visual analysis system for the control strategy of coal-fired power plants, including:

The detailed view module performs an aggregation operation according to the time interval output by the filter view module and the historical operation data, and outputs the control strategy information in the time interval;

The graph view module analyzes and feeds back the propagation mode of the control strategy in space according to the obtained control strategy information;

The strategy view module draws the time cascade corresponding to the control strategy according to the obtained control strategy information;

The visualization module draws the data and information output by the system, and outputs the visualization image information for analyzing the control strategy of the coal-fired power plant.

In order to better illustrate the technical effect of the present invention, the actual operation data of a coal-fired power plant is analyzed. The coal-fired power plant is equipped with 203 sensors, including 158 control sensors and 45 state sensors. Set the first derivative threshold to 80% of the absolute maximum value.

As shown in Figure 2, an incomplete control strategy is specified. It is observed from the historical data that unbalanced combustion occurs in the furnace, and the temperature on the right side is higher than that on the left side. In order to reduce the harmful effects of unbalanced combustion, the desuperheating water valve on the right side should be adjusted.

However, the subsequent impact of this adjustment and the time delay in propagating to power generation efficiency is unclear, so this control strategy is specified in the input panel of the filter view, as shown in the figure. After the matching results are displayed in the sorting panel, because the first result has the highest matching score and the power generation efficiency has changed significantly within the time interval, this result is selected for further analysis.

Exploring the spatial spread of control policy effects: To see relationships between sensors more clearly, select the relational schema. Click on the highlighted sensor in the desuperheated water to analyze the propagation link of the impact. It can be observed that the effect propagates from the desuperheating water to the superheater, to the carbon content of the fly ash, and finally to the power generation efficiency. Among them, the desuperheater and superheater, the carbon content of fly ash and the power generation efficiency are all negatively correlated.

Explore the cascading relationship of control strategy effects in time: First, align the histograms by time delays to observe the cascading effects and time delays of control strategies. It can be seen that 1 minute after the unbalanced combustion occurs, the desuperheating water on the right is increased; after another 5 minutes, the temperature of the superheater is reduced; finally, the power generation efficiency is improved. Based on the above observations and domain knowledge, it can be deduced that adjusting the desuperheating water on the right side can avoid the harmful effects of unbalanced combustion from spreading;

Next, edit the control strategy to understand the cause of unbalanced combustion. Expand the furnace temperature node forward to obtain a complete control strategy. Multiple dampers are tuned up synergistically before unbalanced burn can be seen, which could be the cause of unbalanced burn.

Check and verify conclusions: Check the details of the relevant sensor in the detail view. First, check the temperature difference between the left and right sides of the furnace 1st stage and the furnace 2nd stage, and subtract the left side temperature from the right side temperature. It can be seen that the temperature on the right side is obviously higher, and the unbalanced combustion phenomenon is obvious. Then, looking at the data synthesis for the co-tuning damper, you can see the aggregated characteristics of the co-tuning. It can be inferred that the adjustment of the damper caused the unbalanced combustion.

Finally, import this control strategy into the filter view, and you can see that the support value reaches more than 75%. Therefore, it can be confirmed that timely adjustment of the desuperheating water valve can effectively reduce the negative impact of unbalanced combustion.

As shown in Figure 3, it is the output content of a separate filter view module: it is used to display the historical status of coal-fired power plants and help determine the abnormal time interval, or assist in querying and extracting control strategies, and filter to obtain the time with specific control strategies Interval; filter view has a line chart designed for forward analysis tasks, and an input panel and sorting panel for backward analysis tasks.

The input panel adopts a multi-column view to assist input section control strategies. Among them, specially designed event icons include important detailed attributes (event trend, sensor type, sensor stage, action or state), icons in the same column mean that events occur simultaneously, and different columns mean that events occur sequentially.

The sorting panel is divided into list mode and group mode, and the default is list mode. After the experts input the specified incomplete control strategies, the system will extract all matching complete control strategies and display them in the list mode for the experts to check and select.

Among them, each row represents a control strategy, and the fan chart shows the matching score. All control strategies are sorted from high to low according to the score; each column corresponds to the sensor event group input panel, showing the matching of the sensor event group. The light blue bars encode the proportion of matches for this group; the rectangular colored squares show exactly which events were successfully matched. The grouping mode can display an overview of the matching of the input control strategy, which is mainly used in the final verification stage of the system usage process;

The left end shows the number of control strategies in the group, the rectangular colored square still corresponds to the sensor event input panel, and the transparency encodes the proportion of the control strategy in the group matching the event.

In addition, the support between sensor event groups is marked on the connection chain between columns, which can reflect the credibility of event groups.

The line graph shows the time series for key sensors. Key sensors include power generation efficiency, pollution (NOx emissions) and safety (furnace pressure), and the sensors can be switched through the options in the upper right corner. Experts can view data details by zooming and dragging, and select abnormal time intervals by swiping.

As shown in Figure 4, it is the output content of a separate graph view module: it is used to reflect the spatial propagation mode of the influence of the control strategy, and has a context-oriented mode and a relationship-oriented mode; the context-oriented mode is used to display workflow information and actual space Layout, a relationship-oriented pattern is used to simplify the presentation of associations;

The relationship-oriented mode adopts the existing multi-level force-directed graph layout algorithm, and the closer the sensor is, the more relevant it is. To reduce time complexity, this method computes the layout for the hierarchy. The layout algorithm first computes the position of the force-directed graph for all cells, then computes the position of the sensor within each cell, and finally computes the convex hull from the external tangents of the cells belonging to the same component.

The context mode uses a simplified plane structure diagram of coal power plants, because when users pay attention to the actual location of units and workflow information, the abstract node connection diagram is difficult to understand, and users are very familiar with the workflow structure.

As shown in Figure 5, it is the output of a separate policy view module: it is used to depict the time cascade of control policy influence, visualize the topology of the control policy and its lag-aligned time series.

The topology visualization adopts a design based on a node connection graph. Each node represents a sensor, which corresponds to the time information on the right side. The connection of two nodes represents the influence propagation between two sensors.

The higher the node, the greater the time delay, and the influence is propagated from the upper sensor to the lower connected sensor. Here, the thickness of the link also encodes the size of the correlation.

The lag-aligned time series visualization adopts a histogram-based design. Each row of histogram corresponds to the time series data of a sensor. In order to highlight the data change trend, the height of the column and the transparency of the color encode the data value.

Common time series visualizations include line charts, area charts, etc. Histograms are rarely used. The advantage of this design is that each column corresponds to the data of a time step, and the time delay can be seen at a glance.

In addition, the sensors are sorted from top to bottom according to the time delay from large to small. Sensors with the same delay are grouped into one group, and the specific delay value is marked on the lower left of each group.

This view also provides an alignment function, which aligns the sensor data according to the time delay. You can see that the data change interval is concentrated into a column in the vertical direction, and you can more clearly see the impact of influence propagation on the data trend.

To intelligently analyze control strategies, the Strategy View module adds interactions (eg, extend, insert, delete) for editing and drilling down on control strategies.

Since a coal-fired power plant is a complex system with many control strategies running simultaneously, occasionally flawed analysis results may occur, such as possible missing sensor influence relationships in long time delays.

Therefore, this system provides an editing mode to help experts revise the control strategy, and supports three types of interactions: insertion, extension, and deletion. Experts can click the interaction button to edit. Among them, insert is used to insert a new sensor between two connected sensors; extend is used to expand the control strategy and look forward to more associated sensors; delete is used to remove irrelevant or wrong sensors.

As shown in Figure 6, it is the output content of the individual detail view module: it is used to search for sensors and perform aggregation operations (such as averaging, summing, and subtracting) to understand the control strategy more deeply from the raw data.

Among them, the structure diagram on the left shows the spatial position of the sensor components, and the area diagram on the right shows the specific time series.

The left and right endpoints of the area chart are consistent with the histogram of the strategy view. If you choose to align the time series according to the time delay in the strategy view, the time series of the sensors that have appeared in the strategy view will also be aligned according to the time delay in the detail view.

As shown in Figure 7, for the forward query, the input is a known but incomplete control strategy, expressed as a sequence of sensor events sorted by time, and the model needs to find all the control strategies and their subsequent impacts that can be fuzzy matched with it; the algorithm First divide the time series into multiple rising and falling events, then align the events one by one, and then perform fuzzy matching according to the specified conditions, and finally get the query results. Here, the fuzzy matching adopts the existing longest common subsequence algorithm.

The specific process of forward query is as follows:

Step 1-1. Discretize all historical operating data into trend intervals, namely rising intervals, declining intervals, and stable intervals;

Step 1-2. Segment and align all intervals, so that the historical operating data can be expressed as a matrix M _S , where each row represents a sensor, and each column represents a time interval. The values in the matrix indicate the time interval of the sensor within this time interval. The specific trend of , 0 means stable, 1 means decline, 2 means rise;

Steps 1-3, record the control strategy of the forward query as a similar matrix M _T , that is, each row represents the sensors involved in the target control strategy, and each column represents the expected trend change;

Step 1-4, take out the row where the sensor shared with the matrix M _T from the matrix M _S to obtain M _A , make the number of rows of M _A and M _T the same, and then merge each column of the two matrices by row to obtain two a sequence A and T;

Steps 1-5, use the longest common subsequence matching algorithm for sequences A and T, F(i, j) represents the length of the longest common subsequence in the first i items of A and the first j items of T, and the dynamic programming transfer equation as follows:

As shown in Figure 8, for the backward query, the input is a time interval containing the abnormality of the key sensor, which needs to be extended forward from the key sensor to find the sensor events associated with the abnormality, and obtain the complete control strategy leading to the abnormality.

Backward query is divided into five steps:

Step 2-1, add the specified key sensor to the search queue L, set T to 0;

Step 2-2, obtaining the first sensor P in the queue;

Step 2-3. For each sensor Q, shift the time series of Q back by t time steps, and for each t in the range of 1 to 15 minutes, calculate the Pearson coefficient c of P and Q in the specified time interval :

表示所有列队中第一个传感器P的时间序列长度平均值，Y_i+t表示第i+t个传感器的时间序列长度，

表示所有传感器的时间序列长度平均值；In the formula, M represents the time series length, _Xi represents the time series length of the first sensor P in the i-th queue,

Indicates the average time series length of the first sensor P in all queues, Y _i+t indicates the time series length of the i+t sensor,

Indicates the mean value of the time series length of all sensors;

Step 2-4. If the Pearson coefficient c exceeds 0.8, it is considered that P and Q are related, and Q is added to the search queue L, and T is increased by t;

Step 2-5. If the search queue L is empty or T exceeds 2 hours, terminate the search, otherwise return to step 2 to continue.

The complexity of this algorithm is O(N ² M), where N is the number of sensors and M is the length of the time series.

Claims (8)

1. A visual analysis method for the control strategy of a coal-fired power plant, characterized in that it comprises: Step 1. Obtain the historical operation data of the coal-fired power plant, the historical operation data including the equipment data collected by the sensor and the corresponding sensor location data; Step 2. Based on the sensor position data, analyze the device data within a preset time interval to obtain control strategy information within the time interval; Step 3. Import device data, sensor location data and control strategy information into the pre-built visualization model, in the visualization model, draw graphics for the imported data and information, and perform interactive operations on the data in the graphics The output is displayed. 2. The visual analysis method for the control strategy of a coal-fired power plant according to claim 1, wherein the equipment data includes equipment operation status data and control instruction data. 3. The visual analysis method for the control strategy of a coal-fired power plant according to claim 1, wherein the time interval is the range time for selecting sudden changes in power generation efficiency, furnace negative pressure parameters, or pollutant emissions. 4. The visual analysis method for the control strategy of a coal-fired power plant according to claim 1, characterized in that, the analysis in the step 2 adopts an aggregation operation method, for the spatial position where the sensor is located, the distance between adjacent sensors Correlation degree, the relationship between sensor data changes and time series is analyzed. 5. A visual analysis system for a control strategy of a coal-fired power plant, based on the visual analysis method for a control strategy of a coal-fired power plant according to any one of claims 1 to 4, characterized in that it comprises: The filter view module, according to the occurrence of abnormal events, filters and analyzes the historical operation data, and outputs the time interval where the abnormal events are located; The detailed view module performs an aggregation operation according to the time interval output by the filter view module and the historical operation data, and outputs the control strategy information in the time interval; The graph view module outputs the propagation mode of the control strategy in space according to the obtained control strategy information; The strategy view module outputs the time cascade corresponding to the control strategy according to the obtained control strategy information; The algorithm module is used to graphically draw the output results of the filter view module, detail view module, graph view module and strategy view module, and output visual image information for analyzing control strategy problems of coal-fired power plants. 6. The visual analysis system for the control strategy of a coal-fired power plant according to claim 5, wherein the analysis and filtering of the filter view module is based on a time-delay-aware control strategy extraction method, and the control strategy extraction method includes forward lookup and backward lookup. 7. The visual analysis system for the control strategy of a coal-fired power plant according to claim 6, wherein the forward query adopts the longest common subsequence algorithm to divide the abnormal events in the time interval into multiple rising , falling events, fuzzy matching is performed according to preset conditions. 8. The visual analysis system for the control strategy of coal-fired power plants according to claim 6, wherein the backward query performs extended retrieval based on key sensors that feed back abnormal events within a time interval.

Images