A smart procurement decision-making method and decision-making platform based on power big data

Abstract

The present disclosure provides a smart procurement decision method and decision platform based on electric power big data, which uses a target decision analysis algorithm obtained by multi-task training to accurately analyze and obtain the target decision results of the target electric power big data. A first machine learning algorithm is used to extract the refined feature array of the target electric power big data, where the refined feature array is the feature information expression of the system working status of the target electric power big data. In addition, a second machine learning algorithm is used to extract the reference feature array of the target electric power big data, and a third machine learning algorithm is used to extract the associated feature array of the target electric power big data. Here, the reference feature array and the associated feature array are the perfect features of the target electric power big data. In this way, after combining the row arrays of the refined feature array, the reference feature array, and the associated feature array, the decision operator in the target decision analysis algorithm is used to parse the combined target combination array to obtain the decision result of the target electric power big data.

Description

Technical Field

[0001] The present disclosure relates to the field of data processing, and more specifically, to a smart procurement decision method and decision platform based on power big data. Background Art

[0002] With the rapid development and advancement of the power system, the data generated by the power industry is growing exponentially. These data include power supply and demand data, equipment status data, power grid operation data, etc. Traditional analysis methods can no longer meet the needs of processing and analyzing large-scale power data. Therefore, power big data came into being and became an important technical support for the power industry. Power big data is the process of acquiring valuable information and knowledge by collecting, storing and analyzing a large amount of data generated by various sensors, monitoring equipment and information systems in the power system. It can help the power industry achieve intelligent operation management and optimize decision-making. Power big data can be applied to many aspects, such as energy scheduling, supply and demand forecasting, equipment health monitoring, power grid fault diagnosis, etc. For example, by real-time collection and analysis of data such as power market, load demand and renewable energy, the energy scheduling and supply and demand balance of the power system can be optimized, and energy utilization efficiency can be improved. By monitoring equipment status data, temperature, vibration and other data, combined with machine learning and fault diagnosis algorithms, equipment health monitoring, fault warning and advance maintenance can be achieved, and the reliability and efficiency of power equipment can be improved. By analyzing market demand, price fluctuations and other data, it helps power companies formulate reasonable market strategies and decisions and improve market competitiveness. Power big data can include power load data, power generation data, etc. The accuracy of data analysis and decision-making based on power big data needs to be improved to adapt to changing scenarios and needs. Summary of the invention

[0003] In view of this, the embodiments of the present disclosure at least provide a smart procurement decision method and decision platform based on power big data.

[0004] According to one aspect of an embodiment of the present disclosure, a smart procurement decision method based on power big data is provided, which is applied to a smart decision platform, and the method includes:

[0005] Obtain target power big data of the power system to be analyzed, and determine a refined feature array corresponding to the target power big data through a first machine learning algorithm; the first machine learning algorithm belongs to a target decision analysis algorithm associated with the target power big data; the target decision analysis algorithm includes a second machine learning algorithm and a third machine learning algorithm different from the first machine learning algorithm;

[0006] Acquire a reference string set associated with the second machine learning algorithm, and determine a reference feature array corresponding to the target power big data according to the target power big data and the reference strings in the reference string set;

[0007] Acquire a classification decision data element set associated with the third machine learning algorithm, and determine an associated feature array corresponding to the target power big data according to the target power big data and associated data elements in the classification decision data element set;

[0008] The refined feature array, the control feature array and the associated feature array are combined to obtain a target combination array of the target power big data, the target combination array is input into the decision operator of the target decision analysis algorithm, and the decision operator is used to output the target decision result corresponding to the target power big data.

[0009] According to an example of the present disclosure, the step of acquiring target power big data of the power system to be analyzed and determining a refined feature array corresponding to the target power big data by a first machine learning algorithm includes:

[0010] Performing a data splitting operation on the target power big data to obtain split data fields of the target power big data, performing one-hot encoding on the split data fields to obtain a split array representation corresponding to the split data fields;

[0011] Determine the field position of the split data field in the target power big data, perform position embedding on the field position, and obtain a position array representation corresponding to the field position;

[0012] Determine a segmentation array corresponding to the split data field, perform array summation on the split array representation, the position array representation and the segmentation array, and obtain a field array to be refined of the split data field;

[0013] The field array to be refined is input into the first machine learning algorithm in the target decision analysis algorithm, and the field array to be refined is refined using the first machine learning algorithm to obtain a refined field array corresponding to the split data field, and based on the refined field array corresponding to the split data field, a refined feature array corresponding to the target power big data is determined.

[0014] According to an example of the present disclosure, the first machine learning algorithm includes a target feature mining operator; the target feature mining operator includes a multidimensional significance aggregation component, a start normalization component, a forward propagation component, and an end normalization component;

[0015] The step of inputting the field array to be refined into the first machine learning algorithm in the target decision analysis algorithm, refining the field array to be refined using the first machine learning algorithm to obtain a refined field array corresponding to the split data field, and determining a refined feature array corresponding to the target power big data according to the refined field array corresponding to the split data field, comprises:

[0016] In the first machine learning algorithm of the target decision analysis algorithm, the array of fields to be refined is input into the multidimensional significance aggregation component, and feature mining is performed on the array of fields to be refined according to the multidimensional significance aggregation component to obtain a first intermediate layer array associated with the array of fields to be refined;

[0017] Input the field array to be refined and the first intermediate layer array into the initial normalization component, perform jump connection on the field array to be refined and the first intermediate layer array according to the initial normalization component to obtain a first jump array, and normalize the first jump array to obtain a first normalized array corresponding to the field array to be refined;

[0018] Inputting the first normalized array into the forward propagation component, performing feature mining on the first normalized array according to the forward propagation component, and obtaining a second intermediate layer array corresponding to the first normalized array;

[0019] Input the first normalized array and the second intermediate layer array into the end normalization component, perform jump connection on the first normalized array and the second intermediate layer array according to the end normalization component to obtain a second jump array, normalize the second jump array to obtain a second normalized array corresponding to the to-be-refined field array, obtain the refined field array corresponding to the split data field based on the second normalized array, and determine the refined feature array corresponding to the target power big data based on the refined field array corresponding to the split data field.

[0020] According to an example of the present disclosure, the multi-dimensional saliency aggregation component includes a selected attention component, a starting densely connected component corresponding to the selected attention component, an array combination component, and an ending densely connected component;

[0021] The array combination component is used to combine the feature vectors output by each attention component in the multi-dimensional saliency aggregation component into an array;

[0022] One attention component corresponds to one starting densely connected component;

[0023] In the first machine learning algorithm of the target decision analysis algorithm, the array of fields to be refined is input into the multidimensional significance aggregation component, and feature mining is performed on the array of fields to be refined according to the multidimensional significance aggregation component to obtain a first intermediate layer array associated with the array of fields to be refined, including:

[0024] In a first machine learning algorithm of the target decision analysis algorithm, obtaining a selected attention component from a plurality of attention components possessed by the multidimensional saliency aggregation component;

[0025] Determine a query matrix, a key matrix, and a value matrix associated with the array of fields to be refined according to the array of fields to be refined and the starting densely connected components corresponding to the selected attention components;

[0026] Inputting the query matrix, the key matrix and the value matrix into the selected attention component, processing the query matrix, the key matrix and the value matrix according to the selected attention component, and obtaining an output array corresponding to the selected attention component;

[0027] When each attention component in the multi-dimensional saliency aggregation component is determined as the selected attention component, an output array corresponding to each attention component is obtained, and the output array corresponding to each attention component is array-combined by the array combination component to obtain an attention combination array associated with the array of fields to be refined;

[0028] The focus combination array is input into the end densely connected component, and array feature mining is performed on the focus combination array according to the end densely connected component to obtain a first intermediate layer array associated with the to-be-refined field array.

[0029] According to an example of the present disclosure, the acquiring of a control string set associated with the second machine learning algorithm, and determining a control feature array corresponding to the target power big data according to the target power big data and the control strings in the control string set, include:

[0030] Acquire a comparison string set associated with the second machine learning algorithm, compare the target power big data with the comparison strings in the comparison string set, and obtain a comparison result associated with the target power big data;

[0031] If the comparison result indicates that the comparison string set contains a comparison string corresponding to the target power big data, the comparison string corresponding to the target power big data is determined as the target comparison string;

[0032] Inputting the field chain corresponding to the target control string into the second machine learning algorithm, performing vector conversion on the field chain according to the second machine learning algorithm, and obtaining a control field array corresponding to the target control string;

[0033] According to the comparison field array, a comparison feature array corresponding to the target power big data is determined.

[0034] According to an example of the present disclosure, the method further includes:

[0035] If the comparison result indicates that the comparison string set does not contain a comparison string corresponding to the target power big data, a complete comparison array associated with the comparison string set is obtained, and the complete comparison array is used as a comparison feature array corresponding to the target power big data.

[0036] According to an example of the present disclosure, the step of acquiring a classification decision data element set associated with the third machine learning algorithm and determining an associated feature array corresponding to the target power big data according to the target power big data and associated data elements in the classification decision data element set includes:

[0037] Acquire a classification decision data element set associated with the third machine learning algorithm, perform associated data element pairing on the target power big data and associated data elements in the classification decision data element set, and obtain an associated pairing result associated with the target power big data;

[0038] If the association pairing result indicates that the classification decision data element set contains an associated data element corresponding to the target power big data, the associated data element corresponding to the target power big data is used as the target associated data element;

[0039] Inputting the target associated data element into the third machine learning algorithm, performing vector conversion on the target associated data element according to the third machine learning algorithm, and obtaining an associated data element array corresponding to the target associated data element;

[0040] According to the associated data element array, an associated feature array corresponding to the target power big data is determined.

[0041] According to an example of the present disclosure, the method further includes: if the association pairing result represents that the classification decision data element set does not contain associated data elements corresponding to the target power big data, then obtaining a complete association array associated with the classification decision data element set, and using the complete association array as the associated feature array corresponding to the target power big data.

[0042] According to an example of the present disclosure, the tuning process of the decision analysis algorithm includes:

[0043] Obtaining a power big data learning template for tuning a basic decision analysis algorithm and a template decision result of the power big data learning template, and determining a template feature array corresponding to the power big data learning template through a first machine learning algorithm template; the first machine learning algorithm template belongs to a basic decision analysis algorithm associated with the power big data learning template; the basic decision analysis algorithm includes a second machine learning algorithm template and a third machine learning algorithm template that are different from the first machine learning algorithm template;

[0044] Acquire a reference string set associated with the second machine learning algorithm template, and determine a template reference feature array corresponding to the power big data learning template according to the power big data learning template and the reference strings in the reference string set;

[0045] Acquire a classification decision data element set associated with the third machine learning algorithm, and determine a template-associated feature array corresponding to the power big data learning template according to the target power big data and associated data elements in the classification decision data element set;

[0046] According to the template feature array, the template comparison feature array, the template association feature array, the template decision result and the decision operator of the basic decision analysis algorithm, the basic decision analysis algorithm is tuned, and the tuned basic decision analysis algorithm is used as the target decision analysis algorithm.

[0047] According to another aspect of the present disclosure, there is provided a smart decision-making platform, including:

[0048] one or more processors;

[0049] and one or more memories, wherein the memories store computer-readable codes, and when the computer-readable codes are executed by the one or more processors, the one or more processors execute the above method.

[0050] The present disclosure at least includes the following beneficial effects:

[0051] When the target power big data of the power system to be analyzed is obtained, the embodiment of the present disclosure adopts the first machine learning algorithm to determine the extraction feature array corresponding to the target power big data. The first machine learning algorithm belongs to the target decision analysis algorithm associated with the target power big data, and the target decision analysis algorithm also includes the second machine learning algorithm and the third machine learning algorithm which are different from the first machine learning algorithm. After that, the reference string set associated with the second machine learning algorithm is obtained, and the reference feature array corresponding to the target power big data is determined according to the reference string in the target power big data and the reference string set. Then, the classification decision data element set associated with the third machine learning algorithm is obtained, and the associated feature array corresponding to the target power big data is determined according to the associated data element in the target power big data and the classification decision data element set. Then, the extraction feature array, the reference feature array and the associated feature array are array-combined to obtain the target combination array of the target power big data, and the target combination array is input into the decision operator of the target decision analysis algorithm, and the decision operator is used to output the target decision result corresponding to the target power big data. The present disclosure adopts the target decision analysis algorithm obtained by multi-task training to accurately parse and obtain the target decision result of the target power big data. For example, the first machine learning algorithm is used to extract the extraction feature array of the target power big data, and the extraction feature array here is the feature information expression of the system working status of the target power big data. In addition, the second machine learning algorithm is used to extract the reference feature array of the target power big data, and the third machine learning algorithm is used to extract the associated feature array of the target power big data. The reference feature array and the associated feature array here are the perfect features of the target power big data. In this way, after the refined feature array, the reference feature array and the associated feature array (i.e., the feature information expression and perfect features of the system working status) are combined, the decision operator in the target decision analysis algorithm is used to analyze the combined target combination array to obtain the decision result of the target power big data.

[0052] It should be understood that the above general description and the following detailed description are merely exemplary and explanatory, and are not intended to limit the technical solutions of the present disclosure. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] The above and other purposes, features and advantages of the embodiments of the present disclosure will become more apparent by describing the embodiments of the present disclosure in more detail in conjunction with the accompanying drawings. The accompanying drawings are used to provide a further understanding of the embodiments of the present disclosure and constitute a part of the specification. Together with the embodiments of the present disclosure, they are used to explain the present disclosure and do not constitute a limitation of the present disclosure. In the accompanying drawings, the same reference numerals generally represent the same components or steps.

[0054] Figure 1 A schematic diagram of an application scenario provided for an embodiment of the present disclosure.

[0055] Figure 2 A schematic diagram of the implementation flow of a smart procurement decision-making method based on power big data provided in an embodiment of the present disclosure.

[0056] Figure 3 An algorithm architecture of a first machine learning algorithm provided in an embodiment of the present disclosure.

[0057] Figure 4 A schematic diagram of the structure of an intelligent decision-making device provided in an embodiment of the present disclosure.

[0058] Figure 5 A schematic diagram of the hardware entity of a smart decision-making platform provided in an embodiment of the present disclosure. DETAILED DESCRIPTION

[0059] The technical solutions in the embodiments of the present disclosure will be described clearly and completely below in conjunction with the drawings in the embodiments of the present disclosure. Obviously, the described embodiments are only part of the embodiments of the present disclosure, not all of the embodiments. Based on the embodiments in the present disclosure, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present disclosure.

[0060] In order to make the purpose, technical solutions and advantages of the present disclosure clearer, the technical solutions of the present disclosure are further elaborated in detail below in conjunction with the drawings and embodiments. The described embodiments should not be regarded as limiting the present disclosure. All other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of the present disclosure.

[0061] In the following description, reference is made to "some embodiments", which describe a subset of all possible embodiments, but it is understood that "some embodiments" may be the same subset or different subsets of all possible embodiments, and may be combined with each other without conflict. The terms "first / second / third" involved are merely used to distinguish similar objects and do not represent a specific ordering of the objects. It is understood that "first / second / third" may be interchanged in a specific order or sequential order where permitted, so that the embodiments of the present disclosure described herein can be implemented in an order other than that illustrated or described herein.

[0062] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art to which the present disclosure belongs. The terms used herein are only for the purpose of describing the present disclosure and are not intended to limit the present disclosure.

[0063] Figure 1A schematic diagram of an application scenario according to an embodiment of the present disclosure is shown, in which a smart decision-making platform 110 and multiple power data acquisition devices 120 are schematically shown. The smart decision-making platform 110 here can be an independent server for data processing, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, positioning services, and big data and artificial intelligence platforms. The embodiment of the present disclosure does not impose specific restrictions on this. Each of the multiple power data acquisition devices 120 can be a data collection and monitoring device for each link of the power system, such as smart meters and smart instruments.

[0064] The disclosed embodiment provides a smart purchasing decision method based on power big data, which can be executed by a processor of the smart decision platform 110. Figure 2 A schematic diagram of the implementation process of a smart purchasing decision-making method based on power big data provided by the embodiment of the present disclosure is shown in FIG. Figure 2 As shown, the method comprises the following steps:

[0065] STEP 10, obtain the target power big data of the power system to be analyzed, and determine the refined feature array corresponding to the target power big data through the first machine learning algorithm.

[0066] In the disclosed embodiment, the power big data may include a large amount of data generated in each power link of the power system. It mainly includes the following types of data: power generation data, including the operating status, load conditions, power generation capacity, fuel consumption and other data of the generator set; transmission data, including the load conditions, voltage, current, power loss and other data of the transmission line; distribution data, including the load conditions, current, voltage, power factor and other data of the distribution transformer; user data, including the user's power consumption, power consumption habits, user type, power demand forecast and other data; equipment data, including the operating parameters, status, fault records and other data of power equipment (such as transformers, switchgear, electric meters, etc.); weather data, including temperature, humidity, wind speed, weather forecast and other data, which are used to affect load forecasting and stability analysis of the power system. There are many ways to collect power big data, for example, through sensors and monitoring equipment, by installing sensors and monitoring equipment in power generation, transmission, and distribution equipment, to collect parameter data such as current, voltage, temperature, and pressure in real time; using the monitoring and data acquisition system (SCADA), through networking with power equipment, to realize the collection and monitoring of data in all links of the power system; by installing smart meters and smart instruments, to record users' power consumption, power factor and other data in real time, and transmit them to the data collection center; to exchange data with systems in different power links, such as data docking and data sharing with power plants, substations, distribution networks and other systems; to use wireless communication technology and Internet of Things technology to realize data transmission between devices, making data collection more convenient and efficient; to obtain more comprehensive and rich power data resources by sharing and cooperating with other power-related institutions, regulatory departments, and industry partners. Based on this, it is possible to effectively collect and utilize a large amount of data generated in all links of the power system, and provide support for power big data analysis and smart decision-making.

[0067] Afterwards, a data splitting operation is performed on the target power big data to obtain the split data fields of the target power big data. The specific method of data splitting can be to perform a splitting operation based on the split string to split the data into different data fields. For example, use commas to split the data rows in CSV format. The splitting process can adopt time series data splitting, such as determining the splitting window, sliding along the time axis to split the sequence data into fixed time steps, which can be fixed time intervals or dynamically adjusted according to the frequency of the data. At each sliding position, the subsequence in the window is extracted as an independent data sample. After splitting to obtain multiple split data fields, the split data fields are one-hot encoded to obtain the split array representation corresponding to the split data fields. Arrays are common input data for machine learning algorithms. The split array representation is the data vector representation after one-hot encoding, which can be a one-dimensional array (that is, a vector), a two-dimensional array (that is, a matrix), or a higher-dimensional array. Then, the field position of the split data field is determined in the target power big data (i.e., where it is distributed), and the field position is embedded to obtain the position array representation corresponding to the field position. Specifically, the absolute position encoding method such as Sine-Cosine Positional Encoding can be used for position embedding, or relative position encoding can be used for position embedding, and there is no specific limitation. Next, the segmentation array corresponding to the split data field is determined. The segmentation array is an array that distinguishes the data corresponding to different power data types. The power big data is divided into multiple data fragments, each of which has corresponding semantics and context information. Different data fragments are embedded and mapped based on the segmentation array to complete the encoding, so that the algorithm can recognize and utilize the relationship between different data fragments. After that, the split array representation, the position array representation and the segmentation array are array summed (for example, directly added, or added after weighting, that is, the fusion of the three arrays is completed) to obtain the proposed refined field array of the split data field. Input the field array to be refined into the first machine learning algorithm in the target decision analysis algorithm, and use the first machine learning algorithm to refine the field array to be refined (the refining process is also called feature extraction, or encoding), and obtain the refined field array corresponding to the split data field (that is, the field array obtained after the refinement), and determine the refined feature array corresponding to the target power big data according to the refined field array corresponding to the split data field, for example, merge the refined field arrays corresponding to each split data field, and obtain the refined feature array corresponding to the target power big data. Among them, the target decision analysis algorithm associated with the target power big data also includes a second machine learning algorithm and a third machine learning algorithm that are different from the first machine learning algorithm.

[0068] In one implementation, after obtaining the split array representation corresponding to the split data field and the position array representation corresponding to the field position, the split array representation and the position array representation can be directly summed up without using a split array to obtain the proposed refined field array of the split data field.

[0069] In the above implementation, the first machine learning algorithm can be a transformer. The first machine learning algorithm includes a target feature mining operator, which is an algorithm structure for performing the encoding process of the data field. The target feature mining operator includes a multi-dimensional significance aggregation component, a starting normalization component, a forward propagation component and an end normalization component. The multi-dimensional significance aggregation component is an attention component with multiple heads. The normalization component is used to complete the normalization of the data, such as standardization or normalization. The mechanism of the forward propagation component is perceptron processing. In the embodiment of the present disclosure, according to the field array to be refined, the process of obtaining the refined feature array corresponding to the target power big data may include: in the first machine learning algorithm of the target decision analysis algorithm, the field array to be refined is input into the multi-dimensional significance aggregation component, and the field array to be refined is subjected to feature mining according to the multi-dimensional significance aggregation component, and the first intermediate layer array associated with the field array to be refined (that is, the obtained array of the layer between the input and the output) is obtained. The field array to be refined and the first intermediate layer array are input into the starting normalization component, and the field array to be refined and the first intermediate layer array are jump-connected according to the starting normalization component to obtain a first jump array (or residual array), and the first jump array is normalized to obtain a first normalized array corresponding to the field array to be refined. The first normalized array is input into the forward propagation component, and the first normalized array is feature mined according to the forward propagation component to obtain a second intermediate layer array corresponding to the first normalized array. The first normalized array and the second intermediate layer array are input into the final normalization component, and the first normalized array and the second intermediate layer array are jump-connected according to the final normalization component to obtain a second jump array, and the second jump array is normalized to obtain a second normalized array corresponding to the field array to be refined, and the refined field array corresponding to the split data field is obtained according to the second normalized array, and the refined feature array corresponding to the target power big data is determined according to the refined field array corresponding to the split data field.

[0070] It should be noted that, in the first machine learning algorithm, multiple target feature mining operators are usually set, but of course, only one may be set.

[0071] As an example, see Figure 3 , which is an algorithm architecture of a first machine learning algorithm. The first machine learning algorithm includes a target feature mining operator (a target feature mining operator 31 and a target feature mining operator 32, each of which includes multiple Encoders, i.e., encoders), Figure 3In the example, the first machine learning algorithm includes two layers of target feature mining operators, and Va, Vb...Vx are the input data (field array to be refined) of the first machine learning algorithm. Wa, Wb...Wx are the outputs (refined field arrays) of the first machine learning algorithm. The field array to be refined Va, the field array to be refined Vb...the field array to be refined Vx are input into the first machine learning algorithm, and the first machine learning algorithm can be used to output the refined field array Wa, the refined field array Wb...the refined field array Wx. Among them, the refined field array corresponding to the field array to be refined Va can be the refined field array Wa, the refined field array corresponding to the field array to be refined Vb can be the refined field array Wb...the refined field array corresponding to the field array to be refined Vx can be the refined field array Wx. The refined field array Wa, the refined field array Wb...the refined field array Wx are related to the input of the first machine learning algorithm, the field array to be refined is the original array expression of the split data field, and the refined field array is the enhanced array expression after the split data field is fused with the state semantics of the target power big data.

[0072] The proposed refined field array Va, the proposed refined field array Vb, ... the proposed refined field array Vx are input into the target feature mining operator 32 in the first machine learning algorithm, and the refined field array Sa, the refined field array Sb, ... the refined field array Sx can be output according to the target feature mining operator 32. The refined field array corresponding to the proposed refined field array Va can be the refined field array Sa, and the refined field array corresponding to the proposed refined field array Vb can be the refined field array Sb, ... the refined field array corresponding to the proposed refined field array Vx can be the refined field array Sx. Then, the refined field array Sa, the refined field array Sb, ... the refined field array Sx are input into the target feature mining operator 31 in the first machine learning algorithm, and the refined field array Wa, the refined field array Wb, ... the refined field array Wx are output according to the target feature mining operator 31. The refined field array Wa, the refined field array Wb, ... the refined field array Wx are all regarded as the refined field array 33, and the refined feature array corresponding to the target power big data is determined according to the refined field array 33. It should be noted that the output of each target feature mining operator in the first machine learning algorithm is regarded as a refined field array, the refined field array Wa and the refined field array Sa are regarded as the refined field arrays corresponding to the proposed refined field array Va, the refined field array Wb and the refined field array Sb are regarded as the refined field arrays corresponding to the proposed refined field array Vb... The refined field array Wx and the refined field array Sx are regarded as the refined field arrays corresponding to the proposed refined field array Vx.

[0073] In addition, the architecture of the start normalization component and the end normalization component in the target feature mining operator is consistent, for example, both include skip connections and normalization. The forward propagation component in the target feature mining operator is a perceptron, and the dense connection component is a fully connected layer.

[0074] The multi-dimensional saliency aggregation component in the target feature mining operator is constructed by independent attention components (selfattention), and the number of attention components is consistent with the number of attention heads.

[0075] The input of the target feature mining operator is the array of fields to be refined of the split data field, and the output is the array of refined fields corresponding to the split data field. The input of the target feature mining operator (array of fields to be refined) is input into the multidimensional saliency aggregation component, and the first intermediate layer array associated with the array of fields to be refined is obtained according to the multi-head attention mechanism of the multidimensional saliency aggregation component. Then the first intermediate layer array and the array of fields to be refined are input into the residual layer and normalization layer of the starting normalization component, and the first normalized array corresponding to the array of fields to be refined is output according to the starting normalization component. Then the first normalized array is input into the forward propagation component, and the second intermediate layer array corresponding to the first normalized array is output through the perceptron. Finally, the second intermediate layer array and the first normalized array are input into the end normalization component, and the second normalized array corresponding to the array of fields to be refined is output.

[0076] The second normalized array is the refined field array corresponding to the split data field, and the output of the target feature mining operator is the second normalized array. According to the refined field array corresponding to the split data field, the refined feature array corresponding to the target power big data can be obtained, and the output of the target feature mining operator can also be the refined feature array corresponding to the target power big data.

[0077] It should be noted that the multidimensional saliency aggregation component includes a selected attention component, a starting densely connected component corresponding to the selected attention component, an array combination component, and an end densely connected component. The array combination component is used to combine the feature arrays output by each attention component in the multidimensional saliency aggregation component, for example, the arrays are specifically spliced, and one attention component corresponds to one starting densely connected component. Then, according to the field array to be refined, a first intermediate layer array associated with the field array to be refined is obtained, including: in the first machine learning algorithm of the target decision analysis algorithm, a selected attention component is obtained from multiple attention components in the multidimensional saliency aggregation component. According to the field array to be refined and the starting densely connected component corresponding to the selected attention component, a query matrix (query), a key matrix (key), and a value matrix (value) associated with the field array to be refined are determined. The query matrix, the key matrix, and the value matrix are input into the selected attention component, and the query matrix, the key matrix, and the value matrix are processed according to the selected attention component to obtain the output array corresponding to the selected attention component. When each attention component in the multidimensional saliency aggregation component is determined as a selected attention component, the output array corresponding to each attention component is obtained, and the output array corresponding to each attention component is array-combined through the array combination component to obtain an attention combination array associated with the field array to be refined (the array to be paid attention to after splicing). The attention combination array is input into the end dense connection component, and array features of the attention combination array are mined according to the end dense connection component to obtain the first intermediate layer array associated with the field array to be refined. The multidimensional saliency aggregation component may include multiple attention components, and the field array to be refined is linearly adjusted based on the weight matrix in the starting dense connection components corresponding to different attention components. Multiple attention components run in parallel, so that different attention components establish different spaces.

[0078] STEP 20, obtain a reference string set associated with the second machine learning algorithm, and determine a reference feature array corresponding to the target power big data based on the target power big data and the reference strings in the reference string set.

[0079] For example, a reference string set associated with the second machine learning algorithm is obtained, and the reference strings in the target power big data and the reference string set are compared (for example, one-to-one matching) to obtain a reference result associated with the target power big data. If the reference result indicates that the reference string set contains a reference string corresponding to the target power big data, that is, a matching reference string, the reference string corresponding to the target power big data is determined as the target reference string, and the field chain corresponding to the target reference string (that is, the chain structure data composed of multiple data fields corresponding to the reference string, which can also be called a field group) is input into the second machine learning algorithm, and the field chain is vectorized according to the second machine learning algorithm to obtain a reference field array corresponding to the target reference string. According to the reference field array, a reference feature array (for example, a feature vector) corresponding to the target power big data is determined, for example, after the reference field arrays corresponding to each field chain are merged, a reference feature array corresponding to the target power big data is obtained. The reference string is specifically a preset Regex for representing specific knowledge. If the reference result indicates that the reference string set does not contain a reference string corresponding to the target power big data, a perfect reference array associated with the reference string set is obtained, and the perfect reference array is used as the reference feature array corresponding to the target power big data.

[0080] The target power big data may correspond to a plurality of comparison strings, that is, the target power big data may be matched by a plurality of comparison strings.

[0081] According to each control field array in multiple control field arrays, determine the control feature array corresponding to the target power big data, for example, average multiple control field arrays to obtain a control feature array. When determining the control feature array, it can be obtained based on LSTM. Before comparing the target power big data and the control string, the field chain corresponding to the control string in the control string set is vectorized to complete the mapping from field to array, and the control field array corresponding to each control string is obtained, and the control field array corresponding to all control strings in the control string set is saved. In this way, when the target control string corresponding to the target power big data is determined in the control string set, the control field array corresponding to the target control string can be directly obtained based on the saved string.

[0082] STEP30, obtain a classification decision data element set associated with the third machine learning algorithm, and determine the associated feature array corresponding to the target power big data based on the target power big data and the associated data elements in the classification decision data element set.

[0083] The classification decision data element set contains multiple associated data elements (data elements are data fields) for classification decision comparison, obtains the classification decision data element set associated with the third machine learning algorithm, performs associated data element pairing on the target power big data and the associated data elements in the classification decision data element set, and obtains the associated pairing result associated with the target power big data. If the associated pairing result indicates that the classification decision data element set contains associated data elements corresponding to the target power big data, the associated data elements corresponding to the target power big data are used as the target associated data elements. The target associated data elements are input into the third machine learning algorithm, and the target associated data elements are vectorized according to the third machine learning algorithm to obtain an associated data element array corresponding to the target associated data element. According to the associated data element array, the associated feature array corresponding to the target power big data is determined.

[0084] The associated data elements in the disclosed embodiment are extracted from the training set. The control character string and the associated data elements are representative features generated based on prior data of the power system state. The associated data elements are particularly of great guiding significance for abnormal working conditions of the system.

[0085] If the result of the association pairing indicates that the classification decision data element set does not contain the associated data element corresponding to the target power big data, then a complete association array associated with the classification decision data element set is obtained, and the complete association array is used as the associated feature array corresponding to the target power big data. The target power big data may correspond to multiple associated data elements, that is, the target power big data may be matched by multiple associated data elements.

[0086] According to each associated data element array in the multiple associated data element arrays, the associated feature array corresponding to the target power big data is determined, for example, the multiple associated data element arrays are averaged to obtain the associated feature array. The method of determining the associated feature array can be obtained by using LSTM, merging each associated data element array to obtain the associated feature array corresponding to the target power big data.

[0087] Before the target power big data and the associated data elements are paired, the associated data elements in the classification decision data element set are vectorized to obtain the associated data element array corresponding to each associated data element, and the associated data element array corresponding to all the associated data elements in the classification decision data element set is saved. In this way, when the target associated data element corresponding to the target power big data is determined in the classification decision data element set, the associated data element array corresponding to the target associated data element can be directly obtained based on the saved data.

[0088] STEP40, combine the extracted feature array, the reference feature array and the associated feature array to obtain the target combination array of the target power big data, input the target combination array into the decision operator of the target decision analysis algorithm, and use the decision operator to output the target decision result corresponding to the target power big data.

[0089] Among them, the target decision result is used to adjust and guide the power system to be analyzed. For example, the decision result is used to indicate the state of power generation reaching the standard, fault diagnosis and early warning, power demand forecast, etc., and the data preparation and decision result label setting are carried out according to the actual analysis needs. According to different prediction results, provide guidance and reference for subsequent work. Such as energy scheduling, supply and demand forecasting, equipment health monitoring, power grid fault diagnosis, etc. For example, load forecasting and optimization are carried out, based on historical load data and weather factors, the load demand in the future period is predicted, so as to reasonably allocate power resources and optimize energy scheduling. Or, fault diagnosis and maintenance decision-making are carried out, through the analysis of equipment status and sensor data, equipment faults are detected and diagnosed, and decision-making suggestions for repair and maintenance are provided to reduce downtime and maintenance costs. Another example is to conduct energy market analysis and decision-making, through the analysis of market demand, energy price and other data, provide market strategy and decision support for power companies, and improve market competitiveness. Another example is to carry out renewable energy scheduling and integration, use power big data to optimize the scheduling and integration of renewable energy, and improve the utilization rate and supply stability of renewable energy.

[0090] The decision operator is a classifier, such as Softmax. When the target combination array is input into the decision operator of the target decision analysis algorithm, the target power big data is output according to the confidence level of the decision result of the decision operator of the target decision analysis algorithm. The target decision result corresponding to the target power big data is obtained according to the confidence level. Specifically, the confidence level threshold can be set according to actual needs to divide different decision results, for example, greater than the threshold is decision result A, less than the threshold is decision result B, etc. The specific division method and the type of decision result are not limited. In other words, the decision result can be a label to indicate the corresponding decision information, such as information indicating the state of power generation reaching the standard, fault diagnosis and warning, power demand forecast, etc.

[0091] The target decision analysis algorithm obtained by the present disclosure through multi-task training accurately analyzes and obtains the target decision result of the target power big data. For example, the present disclosure extracts the refined feature array of the target power big data through the first machine learning algorithm, and the refined feature array here is the feature information expression of the system working status of the target power big data. At the same time, the present disclosure extracts the reference feature array of the target power big data based on the second machine learning algorithm, and extracts the associated feature array of the target power big data based on the third machine learning algorithm. The reference feature array and the associated feature array are the perfect features (supplementary and auxiliary features) of the target power big data. After the refined feature array, the reference feature array and the associated feature array are array-combined (such as splicing), the combined target combination array is identified based on the decision operator in the target decision analysis algorithm to obtain the corresponding identification label, indicating the decision result of the target power big data.

[0092] In other embodiments, the present application provides a method for optimizing a decision analysis algorithm involved in a smart procurement decision method based on power big data, which specifically includes the following steps:

[0093] STEP100, obtain the power big data learning template and the template decision result of the power big data learning template used to tune the basic decision analysis algorithm, extract the standby data fields of the power big data learning template according to the standby data field generation logic, obtain the standby data field list corresponding to the power big data learning template, and obtain the target standby data field to be input into the property analysis algorithm from the standby data field list.

[0094] For example, obtain the electric power big data learning template (i.e., the training sample) used to tune the basic decision analysis algorithm (the basic decision analysis algorithm is the initial algorithm that needs to be tuned after initialization) and the template decision result of the electric power big data learning template, perform data splitting operation on the electric power big data learning template according to the standby data field generation logic, obtain the template split data field of the electric power big data learning template, splice the template split data field according to the template split data field splicing logic, and obtain the initial standby data field associated with the electric power big data learning template. Obtain the appearance rate of the initial standby data field in the electric power big data learning template, determine the initial standby data field with an appearance rate greater than the preset appearance rate as the intermediate standby data field, determine the sharing parameters between the intermediate standby data field and the template decision result, and use the intermediate standby data field whose sharing parameters meet the preset sharing parameters in the standby data field generation logic as the candidate data field to be screened. According to the number of fields of the template split data field in the candidate data field to be screened, the candidate data field to be screened whose field number is greater than the preset field number is screened from the candidate data field to be screened. A list of candidate data fields corresponding to the electric power big data learning template is generated based on the screened candidate data fields to be screened, and the target candidate data fields to be input into the property parsing algorithm are obtained from the list of candidate data fields. Among them, the shared parameter is the MI value (Mutual Information) between the intermediate candidate data field and the template decision result. The template split data field splicing logic can splice at least one template split data field according to the distribution position of the template split data field in the electric power big data learning template. If there are more than two template split data fields in the initial candidate data field, the order of the template split data fields in the initial candidate data field is consistent with that in the electric power big data learning template. The Skip-gram algorithm can be used to determine the initial candidate data field associated with the electric power big data learning template.

[0095] In this embodiment of the present disclosure, the associated data elements need to meet two conditions. First, the occurrence rate in the power big data learning template is high, and second, the shared parameters are large. The embodiment of the present disclosure determines the initial standby data field that meets the occurrence rate as the intermediate standby data field, and determines the intermediate standby data field that meets the shared parameters as the standby data field to be screened.

[0096] The disclosed embodiment can determine that the number of fields of the initial standby data fields obtained by splicing is greater than the preset number of fields when the template split data fields are spliced ​​according to the template split data field splicing logic. Based on this, when the standby data fields to be screened are determined according to the above two conditions of the associated data element, a list of standby data fields corresponding to the power big data learning template is directly generated according to the standby data fields to be screened, without screening the standby data fields to be screened whose field number is greater than the preset number of fields from the standby data fields to be screened.

[0097] STEP 200 , qualitatively characterize the target candidate data field through a property analysis algorithm to obtain a qualitative result of the target candidate data field, and perform candidate data field verification on the target candidate data field to obtain a candidate data field verification result of the target candidate data field.

[0098] The property parsing algorithm is a machine learning algorithm that has been debugged and is used to perform qualitative analysis (such as score evaluation) on the target candidate data field. The property parsing algorithm can be a transformer.

[0099] STEP300, if the qualitative result indicates that the target candidate data field meets the evaluation requirements in the candidate data field generation logic, and the candidate data field verification result indicates that the target candidate data field meets the verification conditions in the candidate data field generation logic, then the target candidate data field is determined as the first associated data element in the candidate data field list, and the first associated data element is added to the classification decision data element set.

[0100] The classification decision data element set is used to tune the third machine learning algorithm template.

[0101] STEP400, according to the backup data field generation logic, obtain the collaborative power big data that does not exist in the backup data field list, determine the collaborative power data field in the collaborative power big data as the second associated data element, add the second associated data element to the classification decision data element set, and use the first associated data element and the second associated data element in the classification decision data element set as associated data elements in the classification decision data element set.

[0102] Different candidate data fields can point to the same decision result. When a new data field is obtained, if it is consistent with the decision result corresponding to the existing associated data element, it will be matched to a similar associated data element to improve the recall rate.

[0103] For example, through the standby data field generation algorithm, that is, the standby data field generation logic, a standby data field list corresponding to the power big data learning template is obtained, and the standby data fields included in the standby data field list are target standby data fields. Based on the property analysis algorithm, the target standby data field in the standby data field list is evaluated (that is, qualitatively), and the qualitative result of the target standby data field is obtained. The standby data field is checked, for example, manually checked, to obtain the standby data field check result of the target standby data field. The automatically evaluated and manually checked target standby data field is determined as the first associated data element, and the first associated data element is added to the classification decision data element set.

[0104] Among them, when the qualitative result of a certain data field A represents that the data field A does not meet the evaluation requirements in the candidate data field generation logic, and at the same time the candidate data field verification result of a certain data field B represents that the data field B does not meet the verification conditions in the candidate data field generation logic, the other data fields in the candidate data field list are determined as the first related data elements, and the first related data elements are added to the classification decision data element set.

[0105] Among them, the collaborative power data field in the collaborative power big data, that is, the second associated data element, is different from the first associated data element.

[0106] STEP500, determine the template feature array corresponding to the power big data learning template through the first machine learning algorithm template.

[0107] The first machine learning algorithm template belongs to a basic decision analysis algorithm associated with the electric power big data learning template, and the basic decision analysis algorithm also includes a second machine learning algorithm template and a third machine learning algorithm template that are different from the first machine learning algorithm template. The method of determining the template feature array corresponding to the electric power big data learning template based on the first machine learning algorithm template is consistent with the method of determining the refined feature array corresponding to the target electric power big data based on the first machine learning algorithm.

[0108] STEP600, obtain a reference string set associated with the second machine learning algorithm template, and determine a template reference feature array corresponding to the electric power big data learning template based on the electric power big data learning template and the reference strings in the reference string set.

[0109] The method of determining the template comparison feature array corresponding to the power big data learning template based on the second machine learning algorithm template is consistent with the method of determining the comparison feature array corresponding to the target power big data based on the second machine learning algorithm.

[0110] STEP700, obtain a classification decision data element set associated with the third machine learning algorithm, and determine the template-associated feature array corresponding to the power big data learning template based on the target power big data and the associated data elements in the classification decision data element set.

[0111] The method of determining the template associated feature array corresponding to the power big data learning template based on the third machine learning algorithm template is consistent with the method of determining the associated feature array corresponding to the target power big data based on the third machine learning algorithm.

[0112] STEP800, tune the basic decision analysis algorithm according to the template feature array, template comparison feature array, template association feature array, template decision result and the decision operator of the basic decision analysis algorithm, and use the tuned basic decision analysis algorithm as the target decision analysis algorithm.

[0113] For example, the template feature array, the template comparison feature array, and the template association feature array are combined to obtain a template combination array of the power big data learning template, and the template combination array is input into the decision operator of the basic decision analysis algorithm, and the decision operator is used to output the inference decision result to which the power big data learning template belongs. According to the inference decision result and the template decision result, the performance evaluation cost of the basic decision analysis algorithm is determined. When the performance evaluation cost of the basic decision analysis algorithm indicates that the algorithm has not converged, the algorithm parameters of the basic decision analysis algorithm are corrected according to the performance evaluation cost, and the basic decision analysis algorithm after the correction of the algorithm parameters is used as the intermediate decision analysis algorithm, and the intermediate decision analysis algorithm is tuned until the performance evaluation cost of the tuned intermediate decision analysis algorithm indicates that the algorithm converges.

[0114] The template feature array, template comparison feature array and template association feature array are combined, and the cost of the basic decision analysis algorithm is determined according to the template combination array obtained by the array combination and the template decision result of the power big data learning template, and the algorithm parameters of the basic decision analysis algorithm are optimized based on the cost.

[0115] STEP900, obtain the target power big data of the power system to be analyzed, and determine the refined feature array corresponding to the target power big data through the first machine learning algorithm.

[0116] The first machine learning algorithm is a target decision analysis algorithm associated with the target power big data, and the target decision analysis algorithm also includes a second machine learning algorithm and a third machine learning algorithm that are different from the first machine learning algorithm. The first machine learning algorithm and the first machine learning algorithm template can both be called the first algorithm. The first machine learning algorithm and the first machine learning algorithm template are names given to the first algorithm at different stages. The tuning process is called the first machine learning algorithm template, and the application process is called the first machine learning algorithm.

[0117] STEP 1000, obtain a reference string set associated with the second machine learning algorithm, and determine a reference feature array corresponding to the target power big data based on the target power big data and the reference strings in the reference string set.

[0118] The second machine learning algorithm and the second machine learning algorithm template can both be referred to as the second algorithm. The second machine learning algorithm and the second machine learning algorithm template are names of the second algorithm at different stages. When tuning, the second algorithm is the second machine learning algorithm template. When applying, the second algorithm is the second machine learning algorithm. The reference string set associated with the second machine learning algorithm is consistent with the reference string set associated with the second machine learning algorithm template.

[0119] STEP1100, obtain a classification decision data element set associated with the third machine learning algorithm, and determine the associated feature array corresponding to the target power big data based on the target power big data and the associated data elements in the classification decision data element set.

[0120] The third machine learning algorithm and the third machine learning algorithm template can both be referred to as the third algorithm. The third machine learning algorithm and the third machine learning algorithm template are the names of the third algorithm at different stages. When tuning, the third algorithm is the third machine learning algorithm template, and when applied, the third algorithm is the third machine learning algorithm. The set of classification decision data elements associated with the third machine learning algorithm is consistent with the set of classification decision data elements associated with the third machine learning algorithm template.

[0121] STEP1200, combine the extracted feature array, the reference feature array and the associated feature array to obtain the target combination array of the target power big data, input the target combination array into the decision operator of the target decision analysis algorithm, and use the decision operator to output the target decision result corresponding to the target power big data.

[0122] Refer to the following Figure 4 Describe an intelligent decision-making device according to an embodiment of the present disclosure. Figure 4 FIG. 3 shows a schematic diagram of the structure of the intelligent decision-making device 300 according to an embodiment of the present disclosure. Figure 4 As shown, the intelligent decision-making device 300 may include:

[0123] A target data acquisition module 310 is used to acquire target power big data of the power system to be analyzed, and determine a refined feature array corresponding to the target power big data through a first machine learning algorithm; the first machine learning algorithm belongs to a target decision analysis algorithm associated with the target power big data; the target decision analysis algorithm includes a second machine learning algorithm and a third machine learning algorithm different from the first machine learning algorithm;

[0124] A control data acquisition module 320, configured to acquire a control string set associated with the second machine learning algorithm, and determine a control feature array corresponding to the target power big data according to the target power big data and the control strings in the control string set;

[0125] A decision data acquisition module 330 is used to acquire a classification decision data element set associated with the third machine learning algorithm, and determine an associated feature array corresponding to the target power big data according to the target power big data and associated data elements in the classification decision data element set;

[0126] The decision analysis module 340 is used to perform array combination on the refined feature array, the control feature array and the associated feature array to obtain a target combination array of the target power big data, input the target combination array into the decision operator of the target decision analysis algorithm, and use the decision operator to output the target decision result corresponding to the target power big data.

[0127] Since the function of the intelligent decision-making device 300 is similar to that of the above reference Figure 2 The details of the steps of the described smart procurement decision-making method based on power big data are similar, so for the sake of simplicity, the repeated description of some contents is omitted here.

[0128] In addition, the device according to the embodiment of the present disclosure (for example, the smart decision-making platform) can also be used by Figure 5 The exemplary smart decision-making platform architecture shown is implemented. Figure 5 FIG. 1 is a schematic diagram showing the architecture of an exemplary intelligent decision-making platform according to an embodiment of the present disclosure. Figure 5 As shown, the smart decision platform 400 may include a bus 410, one or more CPUs 420, a read-only memory (ROM) 430, a random access memory (RAM) 440, a communication port 450 connected to a network, an input / output component 460, a hard disk 470, etc. The storage device in the smart decision platform 400, such as the ROM 430 or the hard disk 470, can store various data or files used for computer processing and / or communication and program instructions executed by the CPU. Of course, Figure 5 The architecture shown is only exemplary and can be omitted according to actual needs when implementing different devices. Figure 5 One or more components in the smart decision-making platform shown. The device according to the embodiment of the present disclosure can be configured to execute the smart procurement decision-making method based on power big data according to the above-mentioned various embodiments of the present disclosure, or to implement the smart decision-making device according to the above-mentioned various embodiments of the present disclosure.

[0129] The embodiments of the present disclosure may also be implemented as a computer-readable storage medium. Computer-readable instructions are stored on a computer-readable storage medium according to an embodiment of the present disclosure. When the computer-readable instructions are executed by a processor, the smart procurement decision-making method based on power big data according to an embodiment of the present disclosure described with reference to the above figures may be executed. Computer-readable storage media include, but are not limited to, for example, volatile memory and / or non-volatile memory. Volatile memory may include, for example, random access memory (RAM) and / or cache memory (cache), etc. Non-volatile memory may include, for example, read-only memory (ROM), hard disk, flash memory, etc.

[0130] According to an embodiment of the present disclosure, a computer program product or a computer program is also provided, the computer program product or the computer program including computer-readable instructions, the computer-readable instructions being stored in a computer-readable storage medium. The processor of a computer device can read the computer-readable instructions from the computer-readable storage medium, and the processor executes the computer-readable instructions, so that the computer device executes the smart procurement decision-making method based on power big data described in the above-mentioned various embodiments.

[0131] Those skilled in the art will appreciate that the contents disclosed in this disclosure may be subject to various modifications and improvements. For example, the various devices or components described above may be implemented by hardware, or by software, firmware, or a combination of some or all of the three.

[0132] In addition, as shown in the present disclosure and claims, unless the context clearly indicates an exception, the words "a", "an", "a kind" and / or "the" do not specifically refer to the singular, but may also include the plural. The words "first", "second" and similar words used in the present disclosure do not indicate any order, quantity or importance, but are only used to distinguish different components. Similarly, words such as "include" or "comprise" mean that the elements or objects appearing before the word include the elements or objects listed after the word and their equivalents, without excluding other elements or objects. Words such as "connect" or "connected" and similar words are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect.

[0133] In addition, flow charts are used in the present disclosure to illustrate the operations performed by the system according to the embodiments of the present disclosure. It should be understood that the preceding or following operations are not necessarily performed precisely in order. On the contrary, various steps may be processed in reverse order or simultaneously. At the same time, other operations may also be superimposed on these processes, or one or more operations may be removed from these processes.

[0134] Unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the present disclosure belongs. It should also be understood that terms such as those defined in common dictionaries should be interpreted as having a meaning consistent with their meaning in the context of the relevant technology and should not be interpreted in an idealized or highly formal sense, unless explicitly defined as such herein.

[0135] The present disclosure is described in detail above, but it is obvious to those skilled in the art that the present disclosure is not limited to the embodiments described in this specification. The present disclosure can be implemented as a modification and variation without departing from the purpose and scope of the present disclosure as determined by the claims. Therefore, the description in this specification is for the purpose of illustration and does not have any restrictive meaning for the present disclosure.

Claims

1. A smart purchasing decision-making method based on power big data, characterized in that: Applied to the smart decision-making platform, the method includes: Obtain target power big data of the power system to be analyzed, and determine a refined feature array corresponding to the target power big data through a first machine learning algorithm; the first machine learning algorithm belongs to a target decision analysis algorithm associated with the target power big data; the target decision analysis algorithm includes a second machine learning algorithm and a third machine learning algorithm different from the first machine learning algorithm; Acquire a reference string set associated with the second machine learning algorithm, and determine a reference feature array corresponding to the target power big data according to the target power big data and the reference strings in the reference string set; Acquire a classification decision data element set associated with the third machine learning algorithm, and determine an associated feature array corresponding to the target power big data according to the target power big data and associated data elements in the classification decision data element set; Combining the extracted feature array, the reference feature array, and the associated feature array to obtain a target combination array of the target power big data, inputting the target combination array into a decision operator of the target decision analysis algorithm, and using the decision operator to output a target decision result corresponding to the target power big data; The step of acquiring target power big data of the power system to be analyzed and determining a refined feature array corresponding to the target power big data by a first machine learning algorithm includes: Performing a data splitting operation on the target power big data to obtain split data fields of the target power big data, performing one-hot encoding on the split data fields to obtain a split array representation corresponding to the split data fields; Determine the field position of the split data field in the target power big data, perform position embedding on the field position, and obtain a position array representation corresponding to the field position; Determine a segmentation array corresponding to the split data field, perform array summation on the split array representation, the position array representation and the segmentation array, and obtain a field array to be refined of the split data field; Input the field array to be refined into a first machine learning algorithm in the target decision analysis algorithm, use the first machine learning algorithm to refine the field array to be refined, obtain the refined field array corresponding to the split data field, and determine the refined feature array corresponding to the target power big data according to the refined field array corresponding to the split data field; The acquiring of a reference string set associated with the second machine learning algorithm and determining, according to the target power big data and the reference strings in the reference string set, a reference feature array corresponding to the target power big data comprises: Acquire a comparison string set associated with the second machine learning algorithm, compare the target power big data with the comparison strings in the comparison string set, and obtain a comparison result associated with the target power big data; If the comparison result indicates that the comparison string set contains a comparison string corresponding to the target power big data, the comparison string corresponding to the target power big data is determined as the target comparison string; Inputting the field chain corresponding to the target control string into the second machine learning algorithm, performing vector conversion on the field chain according to the second machine learning algorithm, and obtaining a control field array corresponding to the target control string; Determine, according to the reference field array, a reference feature array corresponding to the target power big data; The step of acquiring a classification decision data element set associated with the third machine learning algorithm and determining an associated feature array corresponding to the target power big data according to the target power big data and associated data elements in the classification decision data element set comprises: Acquire a classification decision data element set associated with the third machine learning algorithm, perform associated data element pairing on the target power big data and associated data elements in the classification decision data element set, and obtain an associated pairing result associated with the target power big data; If the association pairing result indicates that the classification decision data element set contains an associated data element corresponding to the target power big data, the associated data element corresponding to the target power big data is used as the target associated data element; Inputting the target associated data element into the third machine learning algorithm, performing vector conversion on the target associated data element according to the third machine learning algorithm, and obtaining an associated data element array corresponding to the target associated data element; According to the associated data element array, an associated feature array corresponding to the target power big data is determined.

2. The method according to claim 1, characterized in that The first machine learning algorithm includes a target feature mining operator; the target feature mining operator includes a multi-dimensional significance aggregation component, a start normalization component, a forward propagation component and an end normalization component; The step of inputting the field array to be refined into the first machine learning algorithm in the target decision analysis algorithm, refining the field array to be refined using the first machine learning algorithm to obtain a refined field array corresponding to the split data field, and determining a refined feature array corresponding to the target power big data according to the refined field array corresponding to the split data field, comprises: In the first machine learning algorithm of the target decision analysis algorithm, the array of fields to be refined is input into the multidimensional significance aggregation component, and feature mining is performed on the array of fields to be refined according to the multidimensional significance aggregation component to obtain a first intermediate layer array associated with the array of fields to be refined; Input the field array to be refined and the first intermediate layer array into the initial normalization component, perform jump connection on the field array to be refined and the first intermediate layer array according to the initial normalization component to obtain a first jump array, and normalize the first jump array to obtain a first normalized array corresponding to the field array to be refined; Inputting the first normalized array into the forward propagation component, performing feature mining on the first normalized array according to the forward propagation component, and obtaining a second intermediate layer array corresponding to the first normalized array; Input the first normalized array and the second intermediate layer array into the end normalization component, perform jump connection on the first normalized array and the second intermediate layer array according to the end normalization component to obtain a second jump array, normalize the second jump array to obtain a second normalized array corresponding to the to-be-refined field array, obtain the refined field array corresponding to the split data field based on the second normalized array, and determine the refined feature array corresponding to the target power big data based on the refined field array corresponding to the split data field.

3. The method according to claim 2, characterized in that The multi-dimensional saliency aggregation component includes a selected attention component, a starting densely connected component corresponding to the selected attention component, an array combination component, and a terminal densely connected component; The array combination component is used to combine the feature vectors output by each attention component in the multi-dimensional saliency aggregation component into an array; One attention component corresponds to one starting densely connected component; In the first machine learning algorithm of the target decision analysis algorithm, the array of fields to be refined is input into the multidimensional significance aggregation component, and feature mining is performed on the array of fields to be refined according to the multidimensional significance aggregation component to obtain a first intermediate layer array associated with the array of fields to be refined, including: In a first machine learning algorithm of the target decision analysis algorithm, obtaining a selected attention component from a plurality of attention components possessed by the multidimensional saliency aggregation component; Determine a query matrix, a key matrix, and a value matrix associated with the array of fields to be refined according to the array of fields to be refined and the starting densely connected components corresponding to the selected attention components; Inputting the query matrix, the key matrix and the value matrix into the selected attention component, processing the query matrix, the key matrix and the value matrix according to the selected attention component, and obtaining an output array corresponding to the selected attention component; When each attention component in the multi-dimensional saliency aggregation component is determined as the selected attention component, an output array corresponding to each attention component is obtained, and the output array corresponding to each attention component is array-combined by the array combination component to obtain an attention combination array associated with the array of fields to be refined; The focus combination array is input into the end densely connected component, and array feature mining is performed on the focus combination array according to the end densely connected component to obtain a first intermediate layer array associated with the to-be-refined field array.

4. The method according to claim 1, characterized in that The method further comprises: If the comparison result indicates that the comparison string set does not contain a comparison string corresponding to the target power big data, a complete comparison array associated with the comparison string set is obtained, and the complete comparison array is used as a comparison feature array corresponding to the target power big data.

5. The method according to claim 1, characterized in that The method also includes: if the association pairing result indicates that the classification decision data element set does not contain any associated data element corresponding to the target power big data, obtaining a complete associated array associated with the classification decision data element set, and using the complete associated array as an associated feature array corresponding to the target power big data.

6. The method according to any one of claims 1 to 5, characterized in that The tuning process of the decision analysis algorithm includes: Obtaining a power big data learning template for tuning a basic decision analysis algorithm and a template decision result of the power big data learning template, and determining a template feature array corresponding to the power big data learning template through a first machine learning algorithm template; the first machine learning algorithm template belongs to a basic decision analysis algorithm associated with the power big data learning template; the basic decision analysis algorithm includes a second machine learning algorithm template and a third machine learning algorithm template that are different from the first machine learning algorithm template; Acquire a reference string set associated with the second machine learning algorithm template, and determine a template reference feature array corresponding to the power big data learning template according to the power big data learning template and the reference strings in the reference string set; Acquire a classification decision data element set associated with the third machine learning algorithm, and determine a template-associated feature array corresponding to the power big data learning template according to the target power big data and associated data elements in the classification decision data element set; According to the template feature array, the template comparison feature array, the template association feature array, the template decision result and the decision operator of the basic decision analysis algorithm, the basic decision analysis algorithm is tuned, and the tuned basic decision analysis algorithm is used as the target decision analysis algorithm.

7. A smart decision-making platform, characterized in that: include: one or more processors; and one or more memories, wherein the memories store computer-readable codes, and when the computer-readable codes are executed by the one or more processors, the one or more processors execute the method as claimed in any one of claims 1 to 6.

Images