A multi-dimensional intelligent detection method for wind power ramping events

Abstract

The present invention discloses a multi-dimensional wind power ramping event intelligent detection method. The present invention is first based on the manual labeling of wind power ramping events under different definitions, integrates the multi-dimensional attribute characteristics of the ramping events under different definitions, and uses the Alexnet neural network model to fully learn the multi-dimensional wind power ramping event characteristics to perform intelligent detection of wind power ramping events; then, based on the confidence level, an uncertainty visual design scheme for the intelligent detection results of the ramping event is designed to help users analyze and judge the detection results of the ramping event from different dimensions, and increase the interpretability analysis of the intelligent detection results of the ramping event. The present invention designs a model dynamic optimization method in combination with human-computer interaction, supports users to interactively explore the accuracy of the detection results, and then interactively iterates and updates samples to achieve dynamic optimization of the intelligent monitoring model. The present invention increases the interpretability scheme of the wind power ramping event detection results to help users interactively explore the uncertainty of the wind power ramping event detection results.

Description

Technical Field

[0001] The present invention belongs to the field of information technology, and in particular relates to a multi-dimensional wind power ramping event intelligent detection method based on an Alexnet neural network model. Background Art

[0002] Wind energy is a clean, renewable resource with significant environmental benefits and is widely integrated into power systems for wind power generation. Wind power generation can effectively mitigate environmental pollution, such as air and water pollution and global warming, while providing a stable power supply for economic growth. However, the intermittent and variable nature of wind power also threatens the safe operation of power systems. Large fluctuations in wind power over short periods of time, such as significant increases or decreases, can lead to wind power ramping events, causing large-scale power outages and other accidents, impacting the reliability, safety, and economic efficiency of power systems, and posing significant safety risks or severe economic losses to power grid operations. The definition of a ramping event varies across different studies and power grids. Accurate detection of wind power ramping events is a crucial prerequisite for studying ramping events, and detection methods based on ramping definitions are widely used in ramping event research. For example, Haiyang et al. defined a ramping event by determining whether the ratio of the wind power difference within a time interval to the interval is greater than a threshold. Kamath et al. also identified wind power ramping events based on the amplitude of the wind power fluctuation within a specific time interval. Different definitions of ramping events have their own advantages, disadvantages and scope of application. Therefore, considering the various characteristics of wind power ramping events under different definitions, a multi-dimensional wind power ramping event intelligent detection method is designed to accurately identify wind power ramping events.

[0003] Visual analysis technology can fully utilize the human visual perception system to intuitively present the detection results of wind turbine ramping events. However, since the wind turbine ramping event detection model can only obtain ramping or non-ramping results and cannot provide interpretation and analysis of the detection results, it is not conducive to further exploration of ramping events. After close discussions with domain experts, they proposed to intuitively present the uncertainty and interpretability of wind turbine ramping event detection results. This can help users to promptly perceive the errors in the ramping event detection results, and then interactively iterate and optimize the detection results, which is conducive to further detection of wind turbine ramping events. Therefore, designing a wind turbine ramping event detection model and visualizing the interpretability of the detection results are very important for wind turbine ramping event detection and model optimization.

[0004] However, there are still three technical challenges in realizing the ideas of domain experts: (1) How to comprehensively consider the various complex characteristics of wind power ramping events and design a wind power ramping event detection model based on machine learning methods; (2) How to design a visualization method to provide interpretability of wind power ramping event detection results, allowing users to perceive the uncertainty of wind power ramping events from different dimensions; and (3) How to iteratively optimize the model based on the wind power ramping event detection results with high uncertainty to improve the accuracy of the detection model. Summary of the Invention

[0005] The purpose of the present invention is to provide a multi-dimensional wind power ramp event intelligent detection method based on the Alexnet neural network model.

[0006] To achieve the above object, the technical solution adopted by the present invention is:

[0007] Step (1). Manually label the data labels of wind power ramping events based on multiple definitions, integrate the multi-dimensional attribute characteristics of wind power ramping events under different definitions, use the Alexnet neural network model to learn the multi-dimensional wind power ramping event characteristics, and then build an intelligent detection model for multi-dimensional wind power ramping events to comprehensively detect wind power ramping events in different scenarios.

[0008] Step (2). Based on the intelligent detection results, a calendar heat map and power curve map are designed to present the detection results of wind power ramping events. Based on the confidence level, interactive bar charts and lollipop charts are designed to visualize the uncertainty of wind power ramping event detection results and increase the interpretability of the detection results.

[0009] Step (3). Combining visualization with human-computer interaction, design scatter plots, parallel coordinate plots, and line graphs of events with similar power trends to assist users in interactively judging the accuracy of the selected event detection results. For events with incorrect detection, the control panel supports batch iterative updates of sample labels, thereby achieving dynamic optimization of the intelligent model for wind power ramping events.

[0010] Furthermore, in step (1), intelligent detection of wind power ramp events is performed according to the following method:

[0011] (1-1) The user sets the time interval as an event, and manually labels the dataset based on the wind power ramping event characteristics defined by the following five definitions to obtain a labeled training set;

[0012] The user pre-sets five threshold parameters λ1, λ2, λ3, λ4, and λ5;

[0013] Definition 1. If |P(t+ΔT)-P(t)| > λ1, a ramp event occurs; otherwise, no ramp event occurs. ΔT is the set time interval, P(t) is the wind power at time t, and P(t+ΔT) is the wind power at time t+ΔT.

[0014] Definition 2. If max(P[t,t+ΔT])-min(P[t,t+ΔT])>λ2, a ramping event occurs; otherwise, no ramping event occurs. P[t,t+ΔT] is the wind power sequence from time t to time t+ΔT.

[0015] Definition 3. If Then a climbing event occurs, otherwise no climbing event occurs;

[0016] Definition 4. If A ramp event occurs, otherwise no ramp event occurs; the wind power filter signal at time t H represents the maximum value of the time window, P t+h is the wind power at time t+h, P t+h-H is the wind power at time t+hH;

[0017] Definition 5. If A ramp event occurs, otherwise no ramp event occurs; the change in wind power of the nth sample in the wind power sequence is N is the number of samples in the wind power sequence within the set time interval, t n Indicates the time of the nth sample, P(t n+1 ) represents the wind power of the n+1th sample in the wind power sequence, P(t n-1 ) represents the wind power of the n-1th sample in the wind power sequence.

[0018] (1-2) A multi-dimensional wind power ramping event intelligent detection model is constructed based on the Alexnet neural network model. The wind power and type labels of the events in the training set are used as input items of the initial AlexNet network model. The small batch stochastic gradient descent algorithm is used to train the initial AlexNet network model until the preset conditions are met. The target AlexNet network model is obtained to fully learn the characteristics of multi-dimensional wind power ramping events under different definitions and comprehensively detect wind power ramping events in different scenarios.

[0019] (1-3) Multiple power values of the test set data events are used as input values into the multidimensional wind power ramping event intelligent detection model. The multidimensional wind power ramping event feature recognition is performed based on the multidimensional attribute features of the ramping event that the model has learned, and the probability output of the category and the intelligent detection result of whether the ramping event has occurred are obtained.

[0020] (1-4) There are four situations in the intelligent detection results of climbing events: the climbing event actually occurs, and the model detection result is also climbing; the climbing event actually occurs, but the model detection result is no climbing; the climbing event actually does not occur, but the model detection result is that the climbing event occurs; the climbing event actually does not occur, and the model detection result is also no climbing; the number of events in the four situations is N. TP 、N FN 、N FP 、N TN .

[0021] Calculate the accuracy of the detection model performance indicator (accuracy) recall precision Missing report rate and false alarm rate The intelligent detection model of wind power ramping events based on Alexnet neural network and other machine learning-based detection models are evaluated.

[0022] Furthermore, in step (2), the detection results of the wind power ramping event are visualized according to the following method:

[0023] (2-1) Based on the intelligent detection results, a calendar heat map and a power curve map are designed to visualize the detection results of wind power ramping events. The calendar heat map presents the number of wind power ramping events on each day in a calendar format, with colors mapping the number of ramping events. The power curve map shows the power changes at different times of the day, focusing on the power changes when the wind power ramping event occurs, helping users to intuitively perceive the occurrence of the ramping event.

[0024] (2-2) Calculate the confidence of Alexnet output results based on the output probability of the model to help users perceive the uncertainty of the detection results; the confidence of event climbing Confidence of non-climbing events e is a natural constant, y1 and y2 are the output values of the climbing and non-climbing of the Alexnet model respectively;

[0025] (2-3) Based on the confidence level, interactive bar charts and lollipop charts are designed to visualize the uncertainty of wind power ramp event detection results. The bar charts display the number of events with different certainty ranges. The lollipop charts interact with the bar charts to present the specific confidence levels of the detection results within different uncertainty ranges. This supports users in exploring and analyzing the uncertainty of the detection results and increases the interpretability of the detection results.

[0026] Furthermore, in step (3), auxiliary verification of the test results and dynamic optimization of the test model are performed according to the following method:

[0027] (3-1) Design a scatter plot to assist in verifying the uncertainty of the ramp event detection results. Use principal component analysis to reduce the dimensionality of multiple power values of the event to a low-dimensional space and map them in the form of scatter points, where each scatter point represents an event, and the color mapping indicates whether a ramp has occurred. Combined with human-computer interaction, select events with low confidence, determine the location of the event in the scatter plot, and use the labels of other events at the same location to assist in verifying the accuracy of the ramp event detection results.

[0028] (3-2) Design a parallel coordinate graph to present the distribution of function values under five traditional definitions, providing auxiliary judgment for the detection results of ramping events; the parallel coordinate axis displays the ramping function values under different definitions, and the gray background displays the function value distribution of all events in the test data under the definition. Events exceeding the defined threshold are ramping events; the parallel coordinate graph combines human-computer interaction to display the function values of different events under the five definitions, so as to judge whether the event has ramped under the five definitions;

[0029] (3-3) Calculate the similarity of the power curves of the selected event and other labeled events n=1,2,…,N, N is the number of power values measured for this event, x n is the nth power value of the selected event, x n ′ is the nth power value of an existing labeled event. The smaller the S value, the more similar the power curves of the two events are. The combination of visualization and human-computer interaction supports users to interactively select events with high uncertainty and assists in judging the accuracy of their event detection results.

[0030] (3-4) Based on the detection error events determined in steps (3-1) to (3-3), human-computer interaction technology is introduced to support batch updating of event labels, and the updated label data is added to the training set. The wind power ramp event detection model is used to re-learn the features and realize model re-training to interactively and dynamically optimize the detection model.

[0031] Beneficial effects of the present invention: The method of the present invention first integrates the multi-dimensional attribute characteristics of the ramping events under different definitions based on the manual labeling of wind power ramping events under different definitions, and uses the Alexnet neural network model to fully learn the multi-dimensional wind power ramping event characteristics, thereby performing intelligent detection of wind power ramping events; then, based on the confidence level, a visual design scheme for the uncertainty of the intelligent detection results of the ramping events is designed to help users analyze and judge the detection results of the ramping events from different dimensions, thereby increasing the interpretability analysis of the intelligent detection results of the ramping events. Furthermore, a dynamic optimization method of the model is designed in combination with human-computer interaction, which supports users to interactively explore the accuracy of the detection results, interactively iterate and update samples, and realize dynamic optimization of the intelligent monitoring model. The method of the present invention realizes an intelligent model for multi-dimensional wind power ramping event detection. While realizing the intelligent detection of multi-dimensional wind power ramping events based on the Alexnet neural network model, it introduces visualization technology to increase the interpretability scheme of the wind power ramping event detection results, helping users to interactively explore the uncertainty of the wind power ramping event detection results, while supporting users to interactively iterate and update the sample labels of detection errors, thereby interactively and dynamically optimizing the intelligent detection model of wind power ramping events. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 It is a schematic diagram of the overall process of the present invention;

[0033] Figure 2 This is a visualization diagram of the wind power ramping event results in the present invention;

[0034] Figure 3 Schematic diagram of auxiliary verification of the climbing detection result in the present invention;

[0035] Figure 4 Schematic diagram of the results of interactive optimization of the model in the present invention. DETAILED DESCRIPTION

[0036] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0037] like Figure 1 As shown in FIG, a multi-dimensional wind power ramping event intelligent detection method based on the Alexnet neural network model has the following specific steps:

[0038] Step (1). Based on manual labeling of wind power ramping event data labels under multiple definitions, the multi-dimensional attribute features of wind power ramping events under different definitions are integrated, and the multi-dimensional wind power ramping event feature learning is performed using the Alexnet neural network model. Then, an intelligent detection model for multi-dimensional wind power ramping events is constructed to comprehensively detect wind power ramping events in different scenarios. The details are as follows:

[0039] (1-1) The user sets a time interval as an event (the default is 30 minutes), and the data set is manually labeled based on the wind power ramping event characteristics defined by the following five definitions to obtain a labeled training set.

[0040] The user pre-sets five threshold parameters λ1, λ2, λ3, λ4, and λ5;

[0041] Definition 1. If |P(t+ΔT)-P(t)| > λ1, a ramp event occurs; otherwise, no ramp event occurs. ΔT is the set time interval, P(t) is the wind power at time t, and P(t+ΔT) is the wind power at time t+ΔT.

[0042] Definition 2. If max(P[t,t+ΔT])-min(P[t,t+ΔT])>λ2, a ramping event occurs; otherwise, no ramping event occurs. P[t,t+ΔT] is the wind power sequence from time t to time t+ΔT.

[0043] Definition 3. If Then a climbing event occurs, otherwise no climbing event occurs;

[0044] Definition 4. If A ramp event occurs, otherwise no ramp event occurs; the wind power filter signal at time t H represents the maximum value of the time window, P t+h is the wind power at time t+h, P t+h-H is the wind power at time t+hH;

[0045] Definition 5. If A ramp event occurs, otherwise no ramp event occurs; the change in wind power of the nth sample in the wind power sequence is N is the number of samples in the wind power sequence within the set time interval, t n Indicates the time of the nth sample, P(t n+1 ) represents the wind power of the n+1th sample in the wind power sequence, P(t n-1 ) represents the wind power of the n-1th sample in the wind power sequence.

[0046] (1-2) A multi-dimensional wind power ramping event intelligent detection model is constructed based on the Alexnet neural network model. The wind power and type labels of the events in the training set are used as input items of the initial AlexNet network model. The small batch stochastic gradient descent algorithm is used to train the initial AlexNet network model until the preset conditions are met, and the target AlexNet network model is obtained to fully learn the characteristics of multi-dimensional wind power ramping events under different definitions and comprehensively detect wind power ramping events in different scenarios.

[0047] (1-3) Multiple power values of the test set data events are used as input values into the multidimensional wind power ramping event intelligent detection model. The multidimensional wind power ramping event feature recognition is performed based on the multidimensional attribute features of the ramping event that the model has learned, and the probability output of the category and the intelligent detection result of whether the ramping event has occurred are obtained.

[0048] (1-4) There are four situations in the intelligent detection results of climbing events: the climbing event actually occurs, and the model detection result is also climbing; the climbing event actually occurs, but the model detection result is no climbing; the climbing event actually does not occur, but the model detection result is that the climbing event occurs; the climbing event actually does not occur, and the model detection result is also no climbing; the number of events in the four situations is N. TP 、N FN 、N FP 、N TN .

[0049] Calculate the accuracy of the detection model performance indicator (accuracy) recall precision Missing report rate and false alarm rate The intelligent detection model of wind power ramping events based on Alexnet neural network and other machine learning-based detection models are evaluated.

[0050] The effectiveness evaluation of the wind power ramping event intelligent detection model performance is as follows:

[0051] The effectiveness of the proposed multi-dimensional wind turbine ramping event intelligent detection method based on the Alexnet neural network model is demonstrated through quantitative comparison. As shown in Table 1, the proposed wind turbine ramping event intelligent detection model based on the Alexnet neural network is compared with other machine learning models: long short-term memory (LSTM) and BP (back propagation) neural networks. The following table compares the detection results of different models:

[0052] Model R A P M F Alexnet 0.987 0.952 0.770 0.001 0.046 LSTM 0.825 0.759 0.871 0.170 0.109 BP 0.676 0.615 0.768 0.247 0.138

[0053] According to the definition of evaluation metrics, a well-performing ramp event detection model must meet the following criteria: accuracy A, recall R, precision P, low false negative rate M, and false positive rate F. As shown in Table 1, the recall R and accuracy A metrics of this model are greater than 95%, both exceeding the evaluation results of the BP and LSTM models. Furthermore, the evaluation results for the low false negative rate M and false positive rate F of this model are lower than those of the LSTM and BP models. Furthermore, the precision P metric is between that of the LSTM and BP models. Therefore, this wind turbine ramp event detection model demonstrates good performance and can be used for wind turbine ramp event detection.

[0054] Step (2). Based on the intelligent detection results, a calendar heat map and power curve map are designed to present the detection results of wind power ramp events. Based on the confidence level, interactive bar charts and lollipop charts are designed to visualize the uncertainty of wind power ramp event detection results and increase the interpretability of the detection results. The details are as follows:

[0055] (2-1) Based on the intelligent detection results, visualization technology is introduced to design a calendar heat map and a power curve map to visualize the detection results of wind power ramping events; the calendar heat map presents the number of wind power ramping events on each day in a calendar format, and the color maps the number of ramping events; the power curve map shows the power changes at different times of a day, focusing on the power changes when the wind power ramping event occurs, to help users intuitively perceive the occurrence of the ramping event.

[0056] (2-2) Calculate the confidence of Alexnet output results based on the output probability of the model to help users perceive the uncertainty of the detection results; the confidence of event climbing Confidence of non-climbing events e is a natural constant, y1 and y2 are the output values of the climbing and non-climbing of the Alexnet model respectively.

[0057] (2-3) Based on the confidence level, interactive bar charts and lollipop charts are designed to visualize the uncertainty of wind power ramp event detection results. The number of events in different certainty ranges is displayed through the bar chart. The lollipop chart interacts with the bar chart to present the specific confidence level of the detection results in different uncertainty ranges. This supports users to explore and analyze the uncertainty of the detection results and increases the interpretability of the detection results.

[0058] Figure 2(c) shows a calendar heat map of the number of wind power ramping events per month in 2016. It can be seen that April 2016 saw a high number of ramping events. (b) shows a power map of the ramping events on April 25, 2016. It can be seen that the ramping events on that day were mainly concentrated between 9:00 AM and 6:00 PM, likely due to rapid changes in wind speed. Furthermore, (a) shows the uncertainty of the detection results on April 25, 2016. Only three events had a certainty between 0.6 and 0.8, while the other detection results had a certainty between 0.9 and 1, indicating that the credibility of these three events is low and will be the subject of further exploration.

[0059] Step (3). Combining visualization with human-computer interaction, designing scatter plots, parallel coordinate plots, and line graphs of events with similar power trends to assist users in interactively judging the accuracy of the selected event detection results. For events with incorrect detection, the control panel supports batch iterative updates of sample labels, thereby achieving dynamic optimization of the intelligent model for wind power ramping events. The details are as follows:

[0060] (3-1) A scatter plot is designed to assist in verifying the uncertainty of the climbing event detection results. The principal component analysis method is used to reduce the multiple power values of the event to a low-dimensional space and map them in the form of scatter points, where each scatter point represents an event and the color mapping indicates whether climbing has occurred. Combined with human-computer interaction, events with low confidence are selected and their position in the scatter plot is determined. The labels of other events at their location are used to assist in verifying the accuracy of the climbing event detection results.

[0061] (3-2) A parallel coordinate graph is designed to present the distribution of function values under five traditional definitions, providing auxiliary judgment for the detection results of climbing events; the parallel coordinate axis displays the climbing function values under different definitions, and the gray background displays the function value distribution of all events in the test data under the definition. Events that exceed the defined threshold are climbing events; the parallel coordinate graph combines human-computer interaction to display the function values of different events under the five definitions to determine whether the event has climbed under the five definitions.

[0062] (3-3) Calculate the similarity of the power curves of the selected event and other labeled events n=1,2,…,N, N is the number of power values measured for this event, x n is the nth power value of the selected event, x n ′ is the nth power value of the existing labeled event. The smaller the S value, the more similar the power curves of the two events are. The combination of visualization and human-computer interaction supports users to interactively select events with high uncertainty and assists in judging the accuracy of their event detection results.

[0063] (3-4) Based on the detection error events determined in steps (3-1) to (3-3), human-computer interaction technology is introduced to support batch updating of event labels, and the updated label data is added to the training set. The wind power ramp event detection model is used to re-learn the features and realize model re-training to interactively and dynamically optimize the detection model.

[0064] Figure 3 The auxiliary verification results for the ramp-up events in the scatter plot are shown. (c1) is the result for the selected event, and (c2)-(c10) are the auxiliary verification results for events with high similarity to the selected event. It can be seen that the detection results for some events with low certainty are inconsistent with the labels of several events with high similarity to the selected event's power curve. This is likely due to model detection errors, providing a basis for further interactive model optimization.

[0065] Figure 4 The following table represents the model optimization results after interactive iterative sample updates. (a1)-(a5) represent events that the model incorrectly detected, and (c1)-(c4) represent the sample label updates for events (a1)-(a4). (b1)-(b5) represent the detection results after the model was retrained and optimized after the sample update. This shows that after interactive iterative sample updates and model retraining, some events that the model incorrectly detected have been updated to correct results. This demonstrates that after interactive model optimization, the model can detect events that it previously failed to detect.

[0066] After using the method of the present invention, it will be more convenient for experts to detect wind power ramping events and explore the characteristics of wind power ramping events. Compared with the traditional wind power ramping event detection model, the present invention increases the uncertainty and interpretability of the detection results. Through the linkage interaction of multiple visualization views, it helps users explore the accuracy of the detection results, supports interactive sample updates, and realizes dynamic optimization of the model.

Claims

1. A multi-dimensional wind power ramping event intelligent detection method, characterized by: Step (1). First, manually label the wind power ramping event data labels based on multiple definitions, and integrate the multi-dimensional attribute characteristics of wind power ramping events under different definitions, as follows: The user sets a time interval as an event, and manually labels the dataset based on the wind power ramping event characteristics defined by the following five definitions to obtain a labeled training set. The user pre-sets five threshold parameters λ1, λ2, λ3, λ4, and λ5; If |P(t+ΔT)-P(t)|>λ1, a ramp event occurs, otherwise no ramp event occurs; ΔT is the set time interval, P(t) is the wind power at time t, and P(t+ΔT) is the wind power at time t+ΔT; If max(P[t,t+ΔT])-min(P[t,t+ΔT])>λ2, a ramp event occurs, otherwise no ramp event occurs; P[t,t+ΔT] is the wind power sequence from time t to time t+ΔT; if Then a climbing event occurs, otherwise no climbing event occurs; if Then a climbing event occurs, otherwise no climbing event occurs; Wind power filter signal at time t h=1,2,…,H, H represents the maximum value of the time window, P t+h is the wind power at time t+h, P t+h-H is the wind power at time t+hH; if A ramp event occurs, otherwise no ramp event occurs; the change in wind power of the nth sample in the wind power sequence is n=1,2,…,N, N is the number of samples in the wind power sequence within the set time interval, t n Indicates the time of the nth sample, P(t n+1 ) represents the wind power of the n+1th sample in the wind power sequence, P(t n-1 ) represents the wind power of the n-1th sample in the wind power sequence; Then, the Alexnet neural network model is used to learn the characteristics of multi-dimensional wind power ramping events and build an intelligent detection model for multi-dimensional wind power ramping events; Step (2). Based on the intelligent detection results, design a calendar heat map and power curve diagram to present the detection results of wind power ramp events; Design interactive bar charts and lollipop charts based on confidence levels to visualize the uncertainty of wind power ramp event detection results and increase the interpretability of detection results; Step (3). Combining visualization with human-computer interaction, designing scatter plots, parallel coordinate plots, and line graphs of events with similar power trends to assist users in interactively judging the accuracy of the selected event detection results; for events with incorrect detection, the control panel supports batch iterative updates of sample labels, thereby achieving dynamic optimization of the intelligent model of wind power ramping events.

2. A multi-dimensional wind power ramp event intelligent detection method according to claim 1, characterized in that Step (1) further includes: (1-1) The wind power and type labels of the events in the training set are used as input items of the initial AlexNet network model, and the small batch stochastic gradient descent algorithm is used to train the initial AlexNet network model until the preset conditions are met to obtain the target AlexNet network model; (1-2) Multiple power values of the test set data events are used as input values into the multi-dimensional wind power ramping event intelligent detection model. The multi-dimensional wind power ramping event feature recognition is performed based on the multi-dimensional attribute features of the ramping event that have been learned by the model, and the probability output of the category and the intelligent detection result of whether the ramping event has occurred are obtained; (1-3) There are four situations in the intelligent detection results of climbing events: the climbing event actually occurs, and the model detection result is also climbing; the climbing event actually occurs, but the model detection result is no climbing; the climbing event actually does not occur, but the model detection result is that the climbing event occurs; the climbing event actually does not occur, and the model detection result is also no climbing; the number of events in the four situations is N. TP 、N FN 、N FP 、N TN ; Calculate the accuracy of the detection model performance metric Recall Accuracy False negative rate and false alarm rate The intelligent detection model of wind power ramping events based on Alexnet neural network and other machine learning-based detection models are evaluated.

3. The multi-dimensional wind power ramping event intelligent detection method according to claim 1, characterized in that: In step (2), the detection results of the wind power ramp event are visualized according to the following method: (2-1) Based on the intelligent detection results, a calendar heat map and power curve map are designed to visualize the detection results of wind power ramping events; The calendar heat map presents the number of wind power ramping events on each day in a calendar format, with colors mapping the number of ramping events; The power curve diagram shows the power changes at different times of a day, highlights the power changes when a wind power ramp event occurs, and helps users intuitively perceive the occurrence of the ramp event; (2-2) Based on the output probability of the model, the confidence level of the Alexnet output result is calculated to help users perceive the uncertainty of the detection results; Confidence of event ramping Confidence of non-climbing events e is a natural constant, y1 and y2 are the output values of the climbing and non-climbing of the Alexnet model respectively; (2-3) Based on the confidence level, interactive bar charts and lollipop charts are designed to visualize the uncertainty of wind power ramp event detection results. The number of events in different certainty ranges is displayed through the bar chart. The lollipop chart interacts with the bar chart to present the specific confidence level of the detection results in different uncertainty ranges. This supports users to explore and analyze the uncertainty of the detection results and increases the interpretability of the detection results.

4. The multi-dimensional wind power ramping event intelligent detection method according to claim 1, characterized in that: In step (3), auxiliary verification of the test results and dynamic optimization of the test model are performed according to the following methods: (3-1) A scatter plot is designed to assist in verifying the uncertainty of the ramp event detection results. The principal component analysis method is used to reduce the multiple power values of the event to a low-dimensional space and map them in the form of scatter plots, where each scatter plot represents an event and the color mapping indicates whether a ramp event occurred. Combined with human-computer interaction, we select low-confidence events, determine their location in the scatter plot, and use the labels of other events at their location to assist in verifying the accuracy of the ramp event detection results. (3-2) Design a parallel coordinate diagram to present the distribution of function values under five traditional definitions, providing auxiliary judgment for the detection results of climbing events; The parallel coordinate axes show the ramp-up function values under different definitions. The gray background shows the function value distribution of all events in the test data under the definition. Events exceeding the defined threshold are ramp-up events. The parallel coordinate diagram is combined with human-computer interaction to display the function values of different events under the five definitions to determine whether the event has a ramp under the five definitions; (3-3) Calculate the similarity of the power curves of the selected event and other labeled events n=1,2,…,N, N is the number of power values measured for this event, x n is the nth power value of the selected event, x n ′ is the nth power value of an existing labeled event. The smaller the S value, the more similar the power curves of the two events are. The combination of visualization and human-computer interaction supports users to interactively select events with high uncertainty and assists in judging the accuracy of their event detection results. (3-4) Based on the detection error events determined in steps (3-1) to (3-3), human-computer interaction technology is introduced to support batch updating of event labels, and the updated label data is added to the training set. The wind power ramp event detection model is used to re-learn the features and realize model re-training to interactively and dynamically optimize the detection model.

Images