Power material demand forecasting method and system based on deep learning

Abstract

The embodiment of the present application provides a method and system for forecasting demand for electric power materials based on deep learning, which uses sample electric power equipment association trajectories and sample material association trajectories as input data, extracts trajectory graph vectors and introduces a first prediction unit and a second prediction unit to generate sample electric power material knowledge vectors and sample power station knowledge vectors, respectively. Furthermore, by calculating the characteristic distance between the two as the target training cost parameter, the basic prediction neural network is finely adjusted and optimized to generate a more accurate target prediction neural network, which can not only effectively capture the complex demand association between power stations and electric power materials, but also automatically adapt to changes in demand patterns in the power system, provide timely and accurate forecasts of electric power material demand for the power system, and provide strong support for material allocation, inventory management, and procurement decisions, effectively reducing inventory costs and improving the overall efficiency and response speed of the supply chain.

Description

Technical Field

[0001] The present invention relates to the field of deep learning technology, and in particular to a method and system for predicting power material demand based on deep learning. Background Art

[0002] In the power industry, effective management and accurate forecasting of power materials are crucial to ensuring the safe and stable operation of the power grid, improving operational efficiency and reducing costs. Traditional power material demand forecasting methods often rely on historical sales data, inventory status and expert experience judgment. These methods are incapable of coping with the complex and changing power market demand. With the continuous expansion of the scale of the power system and the improvement of the level of intelligence, the types and quantities of power materials have increased dramatically, which has put forward higher requirements for the accuracy, real-time and flexibility of demand forecasting.

[0003] Traditional demand forecasting models, such as time series analysis and regression analysis, can reflect historical trends to a certain extent, but they are unable to cope with high-dimensional, nonlinear and dynamically changing power material demand data. In addition, these methods often ignore the complex interaction between power stations and power materials, as well as the uncertainty of this relationship over time, resulting in a large deviation between the forecast results and actual demand. Summary of the invention

[0004] In order to at least overcome the above-mentioned deficiencies in the prior art, an object of the embodiments of the present application is to provide a method and system for predicting power material demand based on deep learning.

[0005] According to one aspect of the present application, a method for predicting power material demand based on deep learning is provided, the method comprising: obtaining X sample power equipment association trajectories and Y sample material association trajectories, each of the sample power equipment association trajectories representing a demand association knowledge point between the sample power station and the sample power material when the sample power station is the demand guiding node, and each of the sample material association trajectories representing a demand association knowledge point between the sample power material and the sample power station when the sample power material is the demand guiding node; extracting a trajectory graph vector of each of the sample power equipment association trajectories and a trajectory graph vector of each of the sample material association trajectories; using the trajectory graph vectors of the Y sample material association trajectories as model loading data of the first prediction unit in the basic prediction neural network to generate a sample power material knowledge vector, and using the sample power material knowledge vector or the first derivative transfer vector in the sample power material knowledge vector. One or X trajectory graph vectors of the sample power equipment associated trajectories are used as model loading data of the second prediction unit in the basic prediction neural network to obtain a sample power station knowledge vector, and the first derivative transfer vector represents a feature vector when the second prediction unit fits the prediction result of the first prediction unit; the feature distance between the sample power material knowledge vector and the sample power station knowledge vector is calculated to obtain a target training cost parameter, and the basic prediction neural network is parameter-learned according to the target training cost parameter to generate a target prediction neural network, and the target prediction neural network is used to predict and output the trajectory graph vectors of the X target power equipment associated trajectories and the trajectory graph vectors of the Y target material associated trajectories, and generate a target power material knowledge vector and a target power station knowledge vector, and the target power material knowledge vector and the target power station knowledge vector are used to determine the target power material demand result.

[0006] According to one aspect of an embodiment of the present application, a deep learning system is provided, which includes a processor and a machine-readable storage medium, wherein the machine-readable storage medium stores machine-executable instructions, and the machine-executable instructions are loaded and executed by the processor to implement the deep learning-based electricity material demand forecasting method in any one of the possible implementations described above.

[0007] According to one aspect of the embodiments of the present application, a computer program product or a computer program is provided, the computer program product or the computer program includes computer instructions, and the computer instructions are stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device executes the methods provided in various optional implementations of the above three aspects.

[0008] In the technical solutions provided in some embodiments of the present application, the embodiments of the present application significantly improve the accuracy and timeliness of the forecast of power material demand by constructing and training an efficient target prediction neural network. Specifically, using the sample power equipment associated trajectory and the sample material associated trajectory as input data, by extracting the trajectory graph vector and introducing the first prediction unit and the second prediction unit, the sample power material knowledge vector and the sample power station knowledge vector are generated respectively. Furthermore, by calculating the characteristic distance between the two as the target training cost parameter, the basic prediction neural network is finely adjusted and optimized to generate a more accurate target prediction neural network, which can not only effectively capture the complex demand association between the power station and the power material, but also automatically adapt to the changes in the demand pattern in the power system, and provide the power system with timely and accurate power material demand forecasts, and provide strong support for material allocation, inventory management and procurement decisions, effectively reducing inventory costs, and improving the overall efficiency and response speed of the supply chain. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required to be used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other relevant drawings can be extracted in combination with these drawings without creative work.

[0010] Figure 1 A flowchart of a method for predicting power material demand based on deep learning provided in an embodiment of the present application;

[0011] Figure 2 A schematic block diagram of the architecture of a deep learning system provided in an embodiment of the present application. DETAILED DESCRIPTION

[0012] The following description is intended to enable one of ordinary skill in the art to implement and combine the present application, and the description is provided in the context of a specific application scenario and its requirements. It will be apparent to one of ordinary skill in the art that various changes may be made to the disclosed embodiments, and that the general principles defined in the present application may be applied to other embodiments and application scenarios without departing from the principles and scope of the present application. Therefore, the present application is not limited to the described embodiments, but should be given the broadest scope consistent with the claims.

[0013] Figure 1 This is a flow chart of a method for predicting power material demand based on deep learning provided by an embodiment of the present application. The method for predicting power material demand based on deep learning is introduced in detail below.

[0014] Step S110, obtaining X sample power equipment association trajectories and Y sample material association trajectories, each of the sample power equipment association trajectories representing a demand association knowledge point between the sample power station and the sample power material when the sample power station is the demand guiding node, and each of the sample material association trajectories representing a demand association knowledge point between the sample power material and the sample power station when the sample power material is the demand guiding node.

[0015] In detail, the sample power equipment association trajectory refers to a set of data points in the power system, which takes a sample power station (such as a thermal power station, a hydropower station, etc.) as the demand guidance node and records the demand association between the sample power station and various power equipment (such as transformers, cables, circuit breakers, etc.). These data points usually include information such as the time, quantity, equipment type, and specifications of the demand, reflecting the actual demand of the sample power station for different power equipment during operation.

[0016] The sample material association trajectory is a set of data points that takes a sample power material (such as insulation material, cable, transformer oil, etc.) as the demand guidance node and records the demand association between the material and different power stations. These data points contain information such as the supply time, quantity, specification, and receiving power station of the material, reflecting the actual demand and usage of the power material by different power stations.

[0017] In this embodiment, for a large power system network, which includes multiple power stations and various power materials, the server first obtains X sample power equipment associated trajectories and Y sample material associated trajectories.

[0018] Taking a specific sample power equipment association trajectory as an example, suppose a sample power station A is a large thermal power station. During its operation, it has a demand for power equipment such as specific models of transformers and high-voltage cables. As time goes by, a series of demand association records related to these power equipment are formed, including information such as the time, quantity, and specifications of the demand, which constitutes a sample power equipment association trajectory with sample power station A as the demand guidance node.

[0019] Another example is a sample material association trajectory. Suppose a certain type of insulation material is used as sample power material B. It is purchased and used by multiple power stations, such as power station C and power station D. These power stations have demand for this material at different times, in different quantities and frequencies, thus forming a sample material association trajectory with sample power material B as the demand guidance node, recording the demand association knowledge points between it and each power station.

[0020] By mining and analyzing the historical data of the entire power system network, the server can obtain a large number of such sample power equipment-related trajectories and sample material-related trajectories, providing a rich data foundation for subsequent analysis and prediction.

[0021] Step S120, extracting the trajectory graph vector of each of the sample power equipment-related trajectories and the trajectory graph vector of each of the sample material-related trajectories.

[0022] In detail, the trajectory graph vector is the process of converting complex associated trajectory data into a numerical vector. For each sample power equipment associated trajectory or sample material associated trajectory, the server will digitally process and encode the multi-dimensional information (such as location, time, quantity, type, etc.) contained therein to form a multi-dimensional numerical vector. This multi-dimensional numerical vector can accurately represent the key information and features in the trajectory, which is convenient for subsequent machine learning models to process and analyze.

[0023] That is, in this embodiment, for each acquired sample power equipment-related trajectory and sample material-related trajectory, the server needs to extract their trajectory graph vectors.

[0024] Taking the sample power equipment associated trajectory as an example, assume that the sample power equipment associated trajectory contains information in multiple dimensions such as the location of the power station, the type of power equipment, the demand time, and the demand quantity. The server will digitize and encode this information to form a multi-dimensional vector, namely the trajectory map vector. For example, the location of the power station can be represented by longitude and latitude coordinates, the type of power equipment can be represented by a specific code, the demand time can be converted to a timestamp, and the demand quantity can be directly used as a numerical value. These numerical values ​​are combined in a certain order to form the trajectory map vector of the sample power equipment associated trajectory.

[0025] For the sample material-related trajectory, it is also assumed that it contains information such as material type, specification, power station supplied, supply time, supply quantity, etc. The server will also convert this information into trajectory vectors in a similar way.

[0026] Through such processing, the server converts complex associated trajectory data into numerical vector form that can be processed and analyzed by machine learning models.

[0027] Step S130, using the trajectory graph vectors of the Y sample material associated trajectories as model loading data of the first prediction unit in the basic prediction neural network to generate a sample power material knowledge vector, and using one of the sample power material knowledge vector or the first derivative transfer vector and the trajectory graph vectors of the X sample power equipment associated trajectories as model loading data of the second prediction unit in the basic prediction neural network to obtain a sample power station knowledge vector, wherein the first derivative transfer vector represents a feature vector when the second prediction unit fits the prediction result of the first prediction unit.

[0028] In detail, the first prediction unit is a component of the basic prediction neural network, which is specifically responsible for processing the trajectory graph vector of the sample material associated trajectory, and may include a first fusion subunit and a first focus allocation subunit. The first fusion subunit is used to extract deep features in the trajectory graph vector and generate knowledge logic graph features; the first focus allocation subunit is used to evaluate the importance of each knowledge logic graph feature and assign a focus attention coefficient to it. Finally, the sample power material knowledge vector is generated through fusion calculation.

[0029] The second prediction unit is also a component of the basic prediction neural network, responsible for processing the sample power material knowledge vector (or the first derived transfer vector) and the trajectory map vector of the sample power equipment association trajectory. Similar to the first prediction unit, the second prediction unit also contains a fusion subunit and a focus allocation subunit to generate sample power plant knowledge vectors. These vectors reflect the deep-level needs and associations between power plants and power equipment.

[0030] The first derivative transfer vector is a special vector that characterizes the feature transfer of the second prediction unit when fitting the prediction result of the first prediction unit. During the model training process, this vector helps to coordinate the outputs between the two prediction units, making the output of the entire prediction neural network more consistent and accurate.

[0031] That is, in this step, the server uses the trajectory graph vector of the sample material associated trajectory as the model loading data of the first prediction unit in the basic prediction neural network to generate the sample power material knowledge vector.

[0032] Assume that there are 100 trajectory graph vectors of sample material-related trajectories. The server sequentially inputs these trajectory graph vectors into the first fusion subunit in the first prediction unit. The first fusion subunit may be a neural network layer based on deep learning, which processes each input trajectory graph vector, extracts deep features therein, and generates knowledge logic graph features of each sample material-related trajectory.

[0033] For example, for a trajectory graph vector of a specific sample material-associated trajectory, the first fusion subunit may extract logical feature information such as the demand distribution pattern and periodicity of the material in different power stations, and encode it into a new vector, namely, the knowledge logic graph feature.

[0034] Then, the server inputs the knowledge logic graph features of the 100 sample material association trajectories into the first focus allocation subunit. The first focus allocation subunit allocates a focus attention coefficient to each knowledge logic graph feature, which indicates the importance of the knowledge logic graph feature in subsequent calculations.

[0035] For example, the knowledge logic graph features of material association trajectories with frequent and large demands may be assigned higher focus attention coefficients, while the knowledge logic graph features of material association trajectories with less or unstable demands may be assigned lower focus attention coefficients.

[0036] Finally, the server performs a fusion calculation on these knowledge logic graph features and the focus attention coefficient according to the first focus allocation subunit. For example, the knowledge logic graph features can be multiplied by the corresponding focus attention coefficient by weighted summation, and then all the results can be added to generate a sample power material knowledge vector.

[0037] On the other hand, the server uses the sample power material knowledge vector or one of the first derived transfer vectors and the trajectory graph vector of the X sample power equipment associated trajectories as model loading data of the second prediction unit in the basic prediction neural network to obtain the sample power plant knowledge vector.

[0038] Assume that there are 50 sample trajectory map vectors of power equipment associated trajectories. The server first integrates the first derived transfer vector with the trajectory map vector of each sample power equipment associated trajectory. For example, the first derived transfer vector can be directly added or concatenated with the trajectory map vector to generate a first converged map vector of each sample power equipment associated trajectory.

[0039] Then, the server inputs the first converged graph vector of each sample power equipment association trajectory into the second fusion subunit. The second fusion subunit processes it and generates the knowledge logic graph features of each sample power equipment association trajectory. For example, it may analyze the logic features such as the power equipment demand association pattern between power stations and the law of equipment replacement in the trajectory graph vector, and encode them into a new vector.

[0040] Next, the server inputs these knowledge logic graph features into the second focus allocation subunit to generate a focus attention coefficient for each sample power equipment associated trajectory.

[0041] Finally, the server performs a fusion calculation on these knowledge logic graph features and the focus attention coefficient according to the second focus allocation subunit to generate a first focus fusion knowledge vector, and integrates it with the first derivative transfer vector again to obtain a sample power station knowledge vector.

[0042] Step S140, calculating the characteristic distance between the sample electric power material knowledge vector and the sample power station knowledge vector, obtaining a target training cost parameter, performing parameter learning on the basic prediction neural network according to the target training cost parameter, generating a target prediction neural network, the target prediction neural network is used to predict and output the trajectory map vectors of the X target electric power equipment associated trajectories and the trajectory map vectors of the Y target material associated trajectories, generating a target electric power material knowledge vector and a target power station knowledge vector, the target electric power material knowledge vector and the target power station knowledge vector are used to determine the target electric power material demand result.

[0043] In detail, the feature distance is a measure of the similarity or difference between two feature vectors. In this example, the feature distance is used to calculate the degree of difference between the sample power material knowledge vector and the sample power plant knowledge vector. By calculating the feature distance, the server can obtain the target training cost parameter, and then adjust and optimize the parameters of the basic prediction neural network.

[0044] The target training cost parameter is a key indicator used to evaluate model performance during neural network training, and is calculated based on the difference (e.g., error) between the model output and the true label. In this example, the target training cost parameter is obtained by calculating the characteristic distance between the sample power material knowledge vector and the sample power plant knowledge vector. This parameter will be used to guide the parameter learning and optimization process of the neural network.

[0045] The target prediction neural network is a trained and optimized neural network model that can predict and output the trajectory map vectors of the newly input X target power equipment associated trajectories and the trajectory map vectors of the Y target material associated trajectories. The target power material knowledge vector and target power station knowledge vector generated by the target prediction neural network can be used to determine the target power material demand results and provide decision support for the material management and dispatching of the power system.

[0046] Therefore, in this embodiment, the server calculates the characteristic distance between the sample power material knowledge vector and the sample power station knowledge vector. This characteristic distance can be calculated by a certain distance measurement method, such as Euclidean distance, cosine distance, etc.

[0047] Assume that the calculated feature distance is D. The server obtains the target training cost parameter based on this feature distance. For example, the feature distance can be directly used as the training cost parameter, or it can be transformed and processed to be more suitable for subsequent neural network parameter learning.

[0048] Then, the server learns the parameters of the basic prediction neural network based on this target training cost parameter. During the parameter learning process, the server adjusts the parameters such as weights and biases in the neural network so that the output of the network can more accurately reflect the true relationship between the sample power equipment associated trajectory and the sample material associated trajectory.

[0049] After multiple iterations and optimizations, the server finally generates a target prediction neural network. This network can predict and output the trajectory map vectors of the newly input X target power equipment associated trajectories and Y target material associated trajectories, and generate target power material knowledge vectors and target power station knowledge vectors.

[0050] For example, in a new power system scenario, a series of new trajectory graph vectors of power equipment-related trajectories and material-related trajectories are input. The target prediction neural network will process and analyze these inputs based on the previously learned patterns and rules to generate corresponding knowledge vectors.

[0051] Finally, the server uses the generated target power material knowledge vector and target power station knowledge vector to determine the target power material demand results. For example, by analyzing the information in these knowledge vectors, the demand quantity, demand time, and demand specifications of different power materials for each power station in the future can be predicted, thereby providing decision support for the procurement, deployment, and management of power materials.

[0052] Based on the above steps, the embodiment of the present application significantly improves the accuracy and timeliness of the power material demand forecast by constructing and training an efficient target prediction neural network. Specifically, using the sample power equipment associated trajectory and the sample material associated trajectory as input data, by extracting the trajectory graph vector and introducing the first prediction unit and the second prediction unit, the sample power material knowledge vector and the sample power station knowledge vector are generated respectively. Furthermore, by calculating the characteristic distance between the two as the target training cost parameter, the basic prediction neural network is finely adjusted and optimized to generate a more accurate target prediction neural network, which can not only effectively capture the complex demand association between power stations and power materials, but also automatically adapt to changes in demand patterns in the power system, and provide the power system with timely and accurate power material demand forecasts, and provide strong support for material allocation, inventory management and procurement decisions, effectively reducing inventory costs, and improving the overall efficiency and response speed of the supply chain.

[0053] In a possible implementation manner, the first prediction unit includes a first fusion subunit and a first focus allocation subunit. Step S130 includes:

[0054] Step S131, loading the trajectory graph vector of each sample material association trajectory into the first fusion subunit respectively, generating a knowledge logic graph feature of each sample material association trajectory, wherein the knowledge logic graph feature of each sample material association trajectory represents the logic feature information associated with the material demand in the sample material association trajectory;

[0055] Step S132, using the knowledge logic graph features of the Y sample material association trajectories as model loading data of the first focus allocation subunit, and generating a focus attention coefficient for each sample material association trajectory;

[0056] Step S133, the first focus allocation subunit performs a fusion calculation on the knowledge logic graph features of the Y sample material association trajectories and the focus attention coefficient to generate a sample power material knowledge vector.

[0057] In this embodiment, first, the server obtains 200 sample material-related trajectory vectors, which cover various types of power materials, such as wires, insulators, transformer components, and the like.

[0058] Next, the server loads the trajectory graph vector of each sample material-related trajectory into the first fusion subunit. Taking one of the sample material-related trajectories as an example, it is assumed that this trajectory graph vector contains the relevant information of the material "high-strength insulator", including the power station supplied (such as power station E, power station F, etc.), supply time (such as spring and autumn every year), supply quantity (500 each time), material specifications (high voltage resistance 500 kV), etc.

[0059] After receiving this trajectory graph vector, the first fusion subunit begins to analyze and extract the logical feature information associated with the material demand. For example, it may be found that the demand for this high-strength insulator in power station E will increase in spring because the power station will perform equipment maintenance and upgrades in spring; while the demand for it in power station F is more concentrated in autumn, which may be related to the power peak and equipment maintenance plan in autumn. These logical features are encoded into a new vector, namely the knowledge logic graph feature of the sample material association trajectory of "high-strength insulator".

[0060] The server performed similar processing on all 200 sample material association trajectories and generated 200 knowledge logic graph features.

[0061] Then, the server inputs the knowledge logic graph features of the 200 sample material association trajectories into the first focus allocation subunit, which evaluates the importance of each knowledge logic graph feature and generates a focus attention coefficient for it.

[0062] For example, for the knowledge logic graph feature of "high-strength insulators", if the power stations it supplies are all large key power stations and the demand has strong regularity and stability, then it may be assigned a higher focus coefficient, such as 0.8. For another less commonly used power material, its demand is more scattered and unstable, and the corresponding knowledge logic graph feature may be assigned a lower focus coefficient, such as 0.2.

[0063] The server completes the generation of focus attention coefficients for all 200 knowledge logic graph features.

[0064] Finally, the server integrates and calculates the knowledge logic graph features and corresponding focus attention coefficients of the 200 sample material association trajectories based on the first focus allocation subunit.

[0065] Assume that there are three sample material association trajectories with knowledge logic graph features V1, V2, and V3, and the corresponding focus coefficients C1, C2, and C3, respectively. The server will first multiply V1 by C1, V2 by C2, and V3 by C3, and then add these three products to get a preliminary fusion result.

[0066] The server performs such a weighted sum calculation on the knowledge logic graph features and focus attention coefficients of all 200 sample material association trajectories, and finally generates a sample power material knowledge vector.

[0067] Through such a series of processing, the server extracted valuable knowledge vectors from the complex sample material-related trajectory data, providing an important foundation for subsequent power material demand forecasting and management.

[0068] For example, in subsequent analysis, the server found that the sample power material knowledge vector generated in this way indicated that in the future, the demand for certain specifications and models of power materials may increase significantly due to new power plant construction projects, so preparations for material allocation and procurement should be made in advance.

[0069] In a possible implementation, before step S131, the method further includes: acquiring a second derivative transfer vector, where the second derivative transfer vector represents a feature vector when the first prediction unit fits a prediction result of the second prediction unit.

[0070] Step S131 includes:

[0071] Step S1311, respectively converging the second derived transfer vector with the trajectory graph vector of each of the sample material-associated trajectories to generate a converged graph vector of each of the sample material-associated trajectories;

[0072] Step S1312: Load the converged graph vector of each of the sample material association trajectories into the first fusion subunit respectively to generate a knowledge logic graph feature of each of the sample material association trajectories.

[0073] Before generating the knowledge logic graph feature of each sample material association trajectory, this embodiment needs to first obtain the second derivative transfer vector.

[0074] This second derivative transfer vector is a characteristic vector that characterizes the prediction result of the first prediction unit when fitting the second prediction unit. It is assumed that the second derivative transfer vector contains some macro characteristics of the demand trend of power materials, such as the seasonal variation law of the overall power material demand, the common characteristics of material demand of different types of power stations, etc.

[0075] After the server obtains the second derivative transfer vector, it starts to process the trajectory graph vector of the sample material associated trajectory.

[0076] Assume that there are 100 sample material-related trajectory vectors. Take one of the trajectory vectors as an example, and assume that it corresponds to a certain type of power cable. This trajectory vector contains detailed information such as the power stations supplied by the cable (such as power station G, power station H, etc.), supply time (such as the middle of each month), supply quantity (200 meters each time), cable specifications (voltage withstand 220 kV), etc.

[0077] The server converges the second derived transfer vector with the trajectory graph vector of the sample material-related trajectory. For example, a certain value in the second derived transfer vector indicates that the overall power material demand will increase slightly in summer, and the trajectory graph vector of this specific type of power cable shows that the supply quantity to power station G is large in summer. The relevant elements of these two vectors are added or other specific operations are performed to generate the converged graph vector of the sample material-related trajectory.

[0078] The server performs such a convergence operation on the trajectory graph vectors of all 100 sample material-related trajectories, and obtains 100 converged graph vectors.

[0079] Next, the server loads the converged graph vector of each sample material-associated trajectory into the first fusion subunit respectively.

[0080] Taking one of the converged graph vectors as an example, the first fusion subunit will conduct an in-depth analysis of it. It may find that for this particular model of power cable, the increase in demand for power station G is not only due to the general demand growth in summer, but may also be related to the equipment renewal plan of the power station. The first fusion subunit extracts these complex logical relationships and encodes them into a new vector, which is the knowledge logic graph feature of this sample material association trajectory.

[0081] The server performs similar processing on all 100 aggregated graph vectors and generates 100 knowledge logic graph features.

[0082] By obtaining the second derivative transfer vector and aggregating and further processing it with the trajectory graph vector of the sample material association trajectory, the server can more comprehensively and accurately extract the logical characteristics of the material demand association in each sample material association trajectory, providing more valuable information for subsequent demand forecasting and resource allocation.

[0083] For example, through these processes, the server discovered that due to new energy policy adjustments, the demand for power materials of certain specifications in power stations in certain regions will change significantly. Based on such findings, the server can plan the production and deployment of materials in advance to meet future needs.

[0084] In a possible implementation, step S133 includes:

[0085] Step S1331, performing a fusion calculation on the knowledge logic graph features of the Y sample material association trajectories and the focus attention coefficient according to the first focus allocation subunit to generate a target convergence graph vector;

[0086] Step S1332: Integrate the target convergence graph vector and the second derived transfer vector to generate a sample power material knowledge vector.

[0087] In this embodiment, the server has completed the processing of a large number of sample material association trajectories, and obtained the knowledge logic graph features of Y sample material association trajectories and the corresponding focus coefficients. In this specific scenario, it is assumed that Y is equal to 150.

[0088] First, the server starts to perform fusion calculation on the 150 knowledge logic graph features and focus coefficients based on the first focus allocation subunit. Taking three sample material association trajectories as an example, assuming that the knowledge logic graph feature of the first sample material association trajectory is V1, the corresponding focus coefficient is C1; the second is V2 and C2; the third is V3 and C3.

[0089] The server follows the predetermined fusion calculation rules, such as weighted summation, to first multiply V1 by C1, V2 by C2, and V3 by C3, and then add the three products together, and then perform the same operation on other knowledge logic graph features and focus coefficients. After rounds of calculations, the target convergence graph vector is finally generated.

[0090] It is assumed that this target convergence graph vector contains comprehensive information about the demand distribution of different types of power materials, demand change frequency, etc.

[0091] Next, the server integrates the target convergence graph vector with the second derived transfer vector obtained previously. Assume that the second derived transfer vector contains some higher-level feature information about the overall trend of the power material market and the impact of the macro-economy on power demand.

[0092] The server performs a specific integration operation on the target converged graph vector and the second derivative transfer vector, for example, the corresponding elements of the two vectors may be added together, or the integration may be achieved through a more complex function operation.

[0093] Assume that during the integration process, an element in the target convergence graph vector represents the demand growth trend of a certain type of insulator in the past year, while the corresponding element in the second derivative transfer vector reflects the expected growth trend of the entire insulator market. Through integration, the server can comprehensively consider these two aspects of information and obtain a more comprehensive and accurate sample power material knowledge vector.

[0094] This sample power material knowledge vector not only contains the microscopic features in the associated trajectory of each specific sample material, but also incorporates the macroscopic features related to the entire power material field.

[0095] For example, through such integrated calculations, the server found that due to the development of emerging industries, the demand for high-specification power cables has not only increased significantly in certain specific power stations (this is the micro information derived from the sample material association trajectory), but also has a continuous upward trend from the perspective of the entire industry trend (this is the macro information in the second derivative transfer vector). Based on such sample power material knowledge vectors, the server can provide power companies with more accurate material procurement suggestions, such as increasing the inventory of high-specification power cables in advance to meet possible large-scale demand in the future.

[0096] For another example, for materials corresponding to some power equipment that are about to be eliminated, the server analyzes the sample power material knowledge vector and finds that although some power stations still have a small demand (micro information of the sample material association trajectory), judging from the industry development trend, its demand will gradually decrease (macro information of the second derivative transfer vector), thus suggesting that companies reduce the production and storage of such materials to avoid waste of resources.

[0097] This can provide more valuable decision-making basis for the management and planning of power materials.

[0098] In a possible implementation manner, the second prediction unit includes a second fusion subunit and a second focus allocation subunit. Step S130 further includes:

[0099] Step S134, integrating the first derived transfer vector with the trajectory graph vector of each of the sample power equipment associated trajectories, to generate a first converged graph vector of each of the sample power equipment associated trajectories;

[0100] Step S135, loading the first converged graph vector of each of the sample power equipment associated trajectories into the second fusion subunit respectively, generating a knowledge logic graph feature of each of the sample power equipment associated trajectories, wherein the knowledge logic graph feature of each of the sample power equipment associated trajectories represents the logic feature information of the power equipment associated with the sample power equipment associated trajectories;

[0101] Step S136, using the knowledge logic graph feature of each of the sample power equipment associated trajectories as model loading data of the second focus allocation subunit, and generating a focus attention coefficient of each of the sample power equipment associated trajectories;

[0102] Step S137, performing fusion calculation on the knowledge logic graph features of the X sample power equipment associated trajectories and the focus attention coefficient according to the second focus allocation subunit to generate a first focus fusion knowledge vector;

[0103] Step S138: Integrate the first focused fusion knowledge vector and the first derived transfer vector to generate a sample power station knowledge vector.

[0104] Assume that the server is processing a data analysis task of a large-scale power system, in which multiple sample power equipment associated trajectories and related parameters have been obtained.

[0105] First, the server has the first derived transfer vector, which is a vector containing information about some key features and patterns in the power system. At the same time, the server also obtains the trajectory map vector of X sample power equipment associated trajectories. Assume that X is 80, that is, there are 80 different sample power equipment associated trajectories to be processed.

[0106] Take one of the sample power equipment associated trajectories as an example, assuming that it represents the demand for a certain type of transformer in a power station, and the trajectory graph vector contains detailed information such as the time series of demand, quantity changes, and technical specifications of the transformer. The server integrates the first derived transfer vector with the trajectory graph vector of the sample power equipment associated trajectory.

[0107] For example, an element in the first derived transfer vector represents the preference trend of the overall power system for high-efficiency equipment, while the corresponding element in the trajectory graph vector of the sample power equipment associated trajectory is the energy efficiency requirement of the power station for a specific type of transformer. The server integrates these two vectors through a specific algorithm (such as addition or splicing) to generate the first converged graph vector of the sample power equipment associated trajectory.

[0108] The server performs such an integration operation on all 80 sample power equipment associated trajectories and obtains 80 first converged graph vectors.

[0109] Next, the server loads the first converged graph vector of each sample power equipment associated trajectory into the second fusion subunit respectively.

[0110] Taking one of the first converged graph vectors as an example, the second fusion subunit will conduct an in-depth analysis to extract the logical feature information about the association of power equipment. For example, it may find that the demand for this type of transformer in this power station is related to the update cycle of other related equipment (such as switchgear, cables, etc.), or is closely related to the power station's power generation capacity improvement plan. These logical features are encoded into a new vector, namely the knowledge logic graph feature of this sample power equipment association trajectory.

[0111] The server performs similar processing on all 80 first converged graph vectors and generates 80 knowledge logic graph features.

[0112] Then, the server uses the knowledge logic graph features of the 80 sample power equipment associated trajectories as model loading data for the second focused allocation subunit.

[0113] The second focus allocation subunit generates a focus attention coefficient for each knowledge logic graph feature. Assuming that for a certain knowledge logic graph feature, it reflects the long-term stable demand of a key power station for core power equipment, and this demand has an important impact on the stable operation of the entire power system, then it may be assigned a higher focus attention coefficient. For another knowledge logic graph feature, if the demand it represents is more temporary or minor, it may be assigned a lower focus attention coefficient.

[0114] The server completes the generation of focus attention coefficients for all 80 knowledge logic graph features.

[0115] Next, the server performs a fusion calculation on the knowledge logic graph features and corresponding focus attention coefficients of the 80 sample power equipment association trajectories based on the second focus allocation subunit.

[0116] Assume that there are three sample power equipment associated trajectories with knowledge logic graph features V1, V2, and V3, and the corresponding focus coefficients C1, C2, and C3, respectively. The server will first multiply V1 by C1, V2 by C2, and V3 by C3, and then add these three products to get a preliminary fusion result.

[0117] The server performs such a weighted sum calculation on the knowledge logic graph features and focus attention coefficients of all 80 sample power equipment associated trajectories, and finally generates the first focused fusion knowledge vector.

[0118] Finally, the server integrates this first focused fusion knowledge vector with the previous first derived transfer vector.

[0119] Assume that the first focused fusion knowledge vector contains the specific patterns and rules of each power station's demand for different power equipment, and the first derived transfer vector contains macro information about the technical development trend of power equipment and the overall system upgrade direction. Through a specific integration method (such as addition or other complex function operations), the server generates a sample power station knowledge vector.

[0120] That is, this embodiment can comprehensively consider multiple factors and generate a knowledge vector that accurately reflects the relationship between the power station and the power equipment, providing strong support for the planning, operation and maintenance of the power system.

[0121] For example, by analyzing the knowledge vector of the sample power station, the server found that some old power stations need to speed up the replacement of key power equipment to adapt to the new growth in power demand and technical standards. Or, the server found that due to the increase in the proportion of new energy access to the grid, some power stations need to adjust their power equipment configuration and make plans for the purchase and installation of related equipment in advance.

[0122] This can provide accurate decision-making basis and optimization solutions for the efficient and stable operation of the power system.

[0123] In a possible implementation, step S135 includes:

[0124] Step S1351, extracting a knowledge vector of a target sample power station and a knowledge vector of each associated sample power station having an operational relationship with the target sample power station from a first converged graph vector of a first unit trajectory, wherein the first unit trajectory is one of the X sample power equipment associated trajectories, and the target sample power station is one of the sample power stations in the first unit trajectory;

[0125] Step S1352, performing self-attention processing on the knowledge vector of the target sample power station and the knowledge vector of each of the associated sample power stations respectively, to generate a sample attention coefficient of each of the associated sample power stations in the first unit trajectory;

[0126] Step S1353 , performing fusion calculation on the knowledge vectors corresponding to the associated sample power stations according to the sample attention coefficient of each associated sample power station, and generating a knowledge logic graph feature of the first unit trajectory.

[0127] In this embodiment, it is further assumed that the server is processing a data set including 50 sample power equipment-related trajectories, and is currently processing a first unit trajectory therein.

[0128] In the first converged graph vector of the first unit trajectory, the server first extracts the knowledge vector. Assume that the first unit trajectory involves five power stations, namely, power station M, power station N, power station O, power station P and power station Q, where power station M is determined as the target sample power station.

[0129] The server extracts the knowledge vector of the target sample power station M from the first converged graph vector, which contains detailed information such as the location, scale, historical equipment demand pattern, and current equipment operation status of the power station M. At the same time, the server also extracts the knowledge vectors of the other four associated sample power stations (power stations N, O, P, and Q) that have an operational relationship with the power station M. These knowledge vectors also contain relevant feature information of their respective power stations.

[0130] Next, the server performs self-attention processing on the knowledge vector of the target sample power station M and the knowledge vector of each associated sample power station.

[0131] Taking power station N as an example, the server analyzes the similarity, complementarity, and degree of mutual influence between power stations M and N in terms of equipment requirements through calculation and comparison. For example, it is found that power station M has recently carried out a large-scale equipment upgrade, and power station N may also have similar equipment upgrade requirements in the near future because it is in the same power supply area as power station M and has similar load requirements. Through such self-attention processing, the server generates a sample attention coefficient for power station N, which indicates the degree of correlation and importance of power station N relative to power station M in terms of equipment requirements.

[0132] The server also performs the same self-attention processing on power stations O, P, and Q, and generates their respective sample attention coefficients.

[0133] Then, the server performs fusion calculation on the knowledge vectors of the corresponding associated sample power stations according to the sample attention coefficient of each associated sample power station.

[0134] Assume that the sample attention coefficient of power station N is 0.7, the sample attention coefficient of power station O is 0.4, the sample attention coefficient of power station P is 0.6, and the sample attention coefficient of power station Q is 0.5. The server multiplies the knowledge vector of power station N by 0.7, the knowledge vector of power station O by 0.4, the knowledge vector of power station P by 0.6, and the knowledge vector of power station Q by 0.5, and then adds these products.

[0135] Through such fusion calculation, the server generates the knowledge logic graph feature of the first unit trajectory, which comprehensively reflects the complex association and logic pattern between the target sample power station M and other related sample power stations in terms of power equipment demand.

[0136] For example, by analyzing the features of the generated knowledge logic graph, the server found that due to the needs of regional economic development, the equipment update of Power Station M as a core power station will drive the synchronous optimization of equipment in surrounding power stations (such as Power Station N and Power Station P). Although the degree of correlation between Power Station O and Power Station Q is slightly lower, they will also be affected to a certain extent indirectly, and may need to prepare for equipment maintenance and spare parts reserves in advance.

[0137] The server performed similar processing on all 50 sample power equipment association trajectories, thereby comprehensively and deeply exploring the inherent logic and association patterns between power equipment demands of various power stations in the power system, providing a solid data foundation and valuable analysis results for subsequent predictions, planning and decision-making.

[0138] In another specific example, suppose the target sample power station is a large power station that mainly supplies power to industrial areas, and its surrounding associated sample power stations include a small power station that supplies power to nearby residential areas and a medium-sized power station that supplies power to commercial centers. Through self-attention processing and fusion calculation, the server finds that when a large power station undergoes major equipment maintenance, the power load of small and medium-sized power stations will be adjusted accordingly, thereby affecting their short-term demand for backup power equipment and power regulation equipment. Based on such knowledge logic graph features, the server can prepare corresponding power supplies for these associated power stations in advance to ensure the stability and reliability of power supply.

[0139] In this way, valuable information can be extracted from complex power equipment-related data, providing strong support for the optimized operation and management of the power system.

[0140] In a possible implementation manner, the method further includes: calculating a first loss function value between the sample power plant knowledge vector and the second derived transfer vector.

[0141] The first derived transfer vector and the second derived transfer vector are adjusted according to the first loss function value.

[0142] And, calculating a second loss function value between the sample power material knowledge vector and the first derived transfer vector.

[0143] The second derived transfer vector and the first derived transfer vector are adjusted according to the second loss function value.

[0144] In this embodiment, a sample power plant knowledge vector and a sample power material knowledge vector have been generated, and a first derived transfer vector and a second derived transfer vector have been obtained.

[0145] First, the server starts to calculate the first loss function value between the sample power plant knowledge vector and the second derived transfer vector. Assume that the sample power plant knowledge vector contains detailed information about the equipment update requirements and power production capacity of multiple power plants, while the second derived transfer vector contains macro-level information about the balance of supply and demand in the power market and the impact of policies and regulations on power plant operations.

[0146] The server uses a specific loss function calculation formula to measure the difference between the two vectors. For example, the mean square error (MeanSquaredError) may be used as the loss function. The server subtracts the corresponding elements of the sample power plant knowledge vector and the second derivative transfer vector, then squares and sums the difference, and finally divides it by the total number of elements to obtain the first loss function value.

[0147] Assuming that the calculated first loss function value is large, which indicates that there is a large deviation between the sample power plant knowledge vector and the second derived transfer vector, the server adjusts the first derived transfer vector and the second derived transfer vector according to the first loss function value.

[0148] The server may use the back propagation algorithm to adjust the two vectors. For example, if the first loss function value is large, the server will fine-tune the element values ​​of the first derivative transfer vector and the second derivative transfer vector according to the gradient information of the network so that they can be closer in subsequent calculations, thereby reducing the loss function value.

[0149] Next, the server calculates the second loss function value between the sample power material knowledge vector and the first derived transfer vector. The sample power material knowledge vector may include information such as the demand quantity, specification requirements, and supply cycle of various power materials, while the first derived transfer vector includes information such as market trends related to power materials and the impact of technological development on material demand.

[0150] The mean square error or other suitable loss function is also used for calculation. Assuming that the value of the second loss function is also large, it means that the matching degree between the sample power material knowledge vector and the first derived transfer vector is not ideal.

[0151] According to the second loss function value, the server makes adjustments again, by adjusting the element values ​​of the first derivative transfer vector and the second derivative transfer vector so that they can better match the sample power material knowledge vector and reduce the loss function value.

[0152] For example, in a specific scenario, the server finds that the power demand of a power station in a certain area has increased significantly due to the entry of a new industrial project, which has led to an urgent need to update the equipment in the sample power station knowledge vector. However, the policy support and market expectations for the region in the second derivative transfer vector do not fully reflect this change, resulting in a large value for the first loss function. The server will adjust the second derivative transfer vector accordingly, increasing the weight of policy and market changes in the region.

[0153] For another example, for a new type of power material, due to its technological advantages, the demand is growing rapidly, but it is not fully reflected in the first derivative transfer vector, making the second loss function value larger. The server will adjust the first derivative transfer vector to increase the attention to this new type of power material.

[0154] By continuously calculating the loss function value and making corresponding adjustments, the server can make the first derivative transfer vector and the second derivative transfer vector better match the sample power plant knowledge vector and the sample power material knowledge vector, thereby improving the accuracy and reliability of the entire prediction model.

[0155] In another practical scenario, the server found that due to seasonal factors, the demand for certain power materials changed periodically, but this periodicity was not accurately captured in the first derivative transfer vector. By calculating the second loss function value, the server was able to discover this deviation and adjust the first derivative transfer vector to more accurately reflect this seasonal demand change.

[0156] Therefore, by calculating the loss function value and making targeted adjustments based on it, the server can continuously optimize the key vectors in the model and provide more accurate support for demand forecasting and resource allocation of the power system.

[0157] In a possible implementation, the second prediction unit includes Z graph autoencoders, a gated selection subunit, and a multi-layer perceptron. Step S130 may also include:

[0158] Step A110, loading the trajectory graph vector of each of the sample power equipment associated trajectories and Z prior template vectors into the Z graph autoencoders, generating Z associated trend vectors of each of the sample power equipment associated trajectories, each of the graph autoencoders generating one associated trend vector, and each of the prior template vectors reflecting the prior attention weight of the corresponding graph autoencoder for the sample power material;

[0159] Step A120, loading the Z associated trend vectors of each of the sample power equipment associated trajectories, the Z prior template vectors and the sample power material knowledge vector into the gate selection subunit, generating a second convergence graph vector of each of the sample power equipment associated trajectories;

[0160] Step A130: Load the second converged graph vectors of the X sample power equipment association trajectories into the multi-layer perceptron to generate a sample power plant knowledge vector.

[0161] In this embodiment, in this scenario, it is assumed that X is 100, that is, there are 100 sample power equipment associated trajectories to be processed, and Z is 5, that is, there are 5 graph autoencoders.

[0162] First, for each sample power equipment associated trajectory, the server loads it into 5 graph autoencoders together with 5 prior template vectors.

[0163] Assume that a sample power equipment association trajectory represents the demand of power station S for a series of power equipment (such as transformers, circuit breakers, switchgear, etc.). The trajectory map vector contains detailed information such as the location, scale, historical equipment failure records, and equipment update plan of power station S.

[0164] The five prior template vectors respectively reflect the prior attention weights of the graph autoencoder for different types of power materials (such as high-voltage equipment, low-voltage equipment, control equipment, etc.).

[0165] The first graph autoencoder may focus on the demand pattern of high-voltage equipment. When it receives the trajectory graph vector and the prior template vector, it will analyze the information related to the high-voltage equipment and generate an associated trend vector. This vector represents a certain trend in the demand for high-voltage equipment of the power station S, such as the demand for a specific type of high-voltage transformer may increase or decrease in the future.

[0166] Similarly, the other four graph autoencoders will generate corresponding correlation trend vectors based on their respective focus and the information received.

[0167] In this way, for each sample power equipment correlation trajectory, the server obtains 5 correlation trend vectors.

[0168] Next, the server loads the five associated trend vectors, five prior template vectors, and the sample power material knowledge vector of each sample power equipment associated trajectory into the gate selection subunit.

[0169] The gated selection subunit generates a second converged graph vector of the associated trajectory of each sample power device according to these inputs.

[0170] For example, for the associated trend vector and the prior template vector of power station S, the gate selection subunit evaluates the importance and relevance of each associated trend vector. If the area where power station S is located is undergoing a large-scale grid upgrade, then the associated trend vector related to the high-voltage equipment update may be given a higher weight.

[0171] At the same time, information about the supply situation and price fluctuations of the overall power material market in the sample power material knowledge vector will also affect the distribution of weights.

[0172] Through such evaluation and calculation, the gate selection subunit generates a second converged graph vector integrating various factors for the sample power equipment associated trajectory of the power station S.

[0173] The server performs the above processing on all 100 sample power equipment associated trajectories and obtains 100 second converged graph vectors.

[0174] Finally, the server loads the second converged graph vector of the 100 sample power equipment associated trajectories into the multi-layer perceptron.

[0175] The multilayer perceptron has multiple hidden layers and can perform complex nonlinear transformation and feature extraction on the input second converged graph vector.

[0176] Assuming that the first hidden layer of the multilayer perceptron receives these second converged graph vectors, it will initially fuse and abstract their features. The second hidden layer may further explore the deep patterns and relationships therein.

[0177] After multiple layers of processing and calculation, the multilayer perceptron finally generates a sample power station knowledge vector.

[0178] This sample power plant knowledge vector integrates the information of 100 sample power equipment association trajectories, and contains rich knowledge about the equipment requirements of each power plant, the relationship with power materials, and the underlying laws of the overall power system operation.

[0179] For example, by analyzing the generated sample power plant knowledge vectors, the server found that due to the application of emerging technologies, the demand for intelligent power equipment in power plants in some areas is showing a rapid growth trend, while power plants in other areas are more concerned about the reliability and stability of equipment, and the demand for maintenance and upgrades of traditional equipment is more prominent.

[0180] Based on such findings, the server can provide power companies with more targeted equipment procurement recommendations, operation and maintenance strategies, and resource allocation plans to improve the operating efficiency and reliability of the power system.

[0181] In another specific scenario, the server finds that the demand pattern of a certain type of power station (such as a wind power station) for specific power equipment (such as inverters and energy storage equipment) is significantly different from that of other types of power stations (such as thermal power stations). The sample power station knowledge vector generated through the above processing steps can clearly reflect this difference, thereby providing strong support for personalized management and services for different types of power stations.

[0182] In summary, through the collaborative work of graph autoencoders, gated selection subunits, and multilayer perceptrons, valuable knowledge vectors can be extracted from complex power equipment associated trajectory data, providing accurate and comprehensive decision-making basis for the optimization and management of power systems.

[0183] In a possible implementation, step A110 includes:

[0184] Step A111, extracting the knowledge vector of the target sample power station and the knowledge vector of each associated sample power station having an operational relationship with the target sample power material from the trajectory graph vector of the second unit trajectory, wherein the second unit trajectory is one of the X sample power equipment associated trajectories, and the target sample power station is one of the sample power materials in the second unit trajectory;

[0185] Step A112, respectively converging the knowledge vector of the target sample power station with the Z prior template vectors to generate Z first converged vectors;

[0186] Step A113, performing self-attention processing on the Z first converged vectors and the knowledge vector of each associated sample power station respectively, to generate a sample attention coefficient of each associated sample power station in the second unit trajectory;

[0187] Step A114 , performing fusion calculation on the knowledge vectors corresponding to the associated sample power stations according to the sample attention coefficient of each associated sample power station, and generating Z associated trend vectors of the second unit trajectory.

[0188] In this embodiment, one of the X sample power equipment-associated trajectories is marked as a second unit trajectory.

[0189] In the trajectory graph vector of the second unit trajectory, the server first extracts the knowledge vector. Assume that the second unit trajectory involves multiple power stations and related power materials, where the target sample power station is determined to be power station T, and there are multiple associated sample power stations that have an operational relationship with the target sample power material, such as power station U, power station V, and power station W.

[0190] The server extracts the knowledge vector of the target sample power station T from the trajectory graph vector of the second unit trajectory, which contains detailed information such as the location, installed capacity, equipment aging, and recent power production data of the power station T. At the same time, the server also extracts the knowledge vectors of each associated sample power station (power station U, V, W) that has an operational relationship with the power station T, and these knowledge vectors contain similar feature information of each power station.

[0191] Next, the server aggregates the knowledge vector of the target sample power station T with the Z prior template vectors. Assuming Z is 3, these three prior template vectors focus on the reliability, energy efficiency and maintenance cost of power equipment respectively.

[0192] The server performs a specific mathematical operation, such as weighted summation, on the knowledge vector of power station T and the first prior template vector (focusing on reliability) to generate the first first converged vector. In the same way, it converges with the second prior template vector (focusing on energy efficiency) and the third prior template vector (focusing on maintenance cost) to generate another two first converged vectors.

[0193] Then, the server performs self-attention processing on the three first converged vectors and the knowledge vector of each associated sample power station.

[0194] Taking power station U as an example, the server analyzes the correlation and importance between the first convergence vector and the knowledge vector of power station U through complex calculations and comparisons. For example, it is found that the reliability requirements of power station T are strongly correlated with those of power station U because they are in the same power grid area and share some transmission lines, thus generating a corresponding sample attention coefficient for power station U.

[0195] Similarly, the second and third first converged vectors are processed with the knowledge vector of the power station U to generate another two sample attention coefficients.

[0196] The same self-attention processing is performed on power station V and power station W to generate corresponding sample attention coefficients for them respectively.

[0197] Finally, the server performs fusion calculation on the knowledge vectors of the corresponding associated sample power stations according to the sample attention coefficient of each associated sample power station.

[0198] Assume that the attention coefficients of the three samples of power station U are 0.6, 0.7 and 0.4 respectively, the attention coefficients of the three samples of power station V are 0.5, 0.8 and 0.3 respectively, and the attention coefficients of the three samples of power station W are 0.7, 0.6 and 0.5 respectively.

[0199] The server multiplies the knowledge vector of power station U by the three sample attention coefficients respectively, and then adds them together to generate an associated trend vector corresponding to the first prior template vector. In the same way, the associated trend vectors corresponding to the second and third prior template vectors are generated.

[0200] The same fusion calculation is performed on power station V and power station W, and finally three associated trend vectors of the second unit trajectory are generated.

[0201] For example, through the above processing, the server finds that since power plant T plans to carry out large-scale equipment upgrades to improve energy efficiency, power plants U and V, which have operational connections with it, also need to adjust their equipment operation strategies accordingly, and the demand for high-efficiency equipment has increased. This discovery is reflected in the generated correlation trend vector, providing valuable information for subsequent analysis and decision-making.

[0202] In another specific scenario, the server found that the maintenance cost of power station T became a key factor, which led to the associated power station W being more inclined to choose equipment with long life and low maintenance requirements. Through the generated correlation trend vector, the server can accurately capture this correlation and trend, providing strong support for the overall planning and resource allocation of the power system.

[0203] In this way, valuable correlation trend information can be mined from complex power equipment correlation trajectory data, providing an important basis for the optimized operation and management of the power system.

[0204] In a possible implementation, step A120 includes:

[0205] Step A121, for each of the sample power equipment associated trajectories, the Z prior template vectors are respectively subjected to self-attention processing with the sample power material knowledge vectors to generate a self-attention coefficient of each of the associated trend vectors in each of the sample power equipment associated trajectories;

[0206] Step A122: for each of the sample power equipment associated trajectories, a fusion calculation is performed on the corresponding Z associated trend vectors according to the self-attention coefficients of the Z associated trend vectors to generate a second converged graph vector for each of the sample power equipment associated trajectories.

[0207] In this embodiment, a specific sample power equipment associated trajectory is taken as an example for detailed description.

[0208] First, for this specific sample power equipment association trajectory, the server has Z association trend vectors, Z prior template vectors, and sample power material knowledge vectors. Assume that Z is 4.

[0209] The server starts self-attention processing. These four prior template vectors represent different focus points, such as the reliability weight, cost weight, performance weight, and environmental protection weight of the equipment.

[0210] The server performs self-attention processing on the first prior template vector and the sample power material knowledge vector. The sample power material knowledge vector contains information such as the type, quantity, quality requirements, and market supply of various power materials. Through complex calculations and comparisons, the server generates the self-attention coefficient of the first associated trend vector in the sample power equipment associated trajectory.

[0211] Similarly, the server performs self-attention processing on the other three prior template vectors and the sample power material knowledge vector, and generates self-attention coefficients of the corresponding second, third and fourth associated trend vectors.

[0212] Next, the server performs fusion calculation on the corresponding four associated trend vectors based on the self-attention coefficients of these four associated trend vectors.

[0213] Assume that the self-attention coefficient of the first associated trend vector is 0.3, the second is 0.2, the third is 0.4, and the fourth is 0.1.

[0214] The server multiplies the first relevance trend vector by 0.3, the second relevance trend vector by 0.2, the third relevance trend vector by 0.4, and the fourth relevance trend vector by 0.1.

[0215] These four products are then added together to obtain a preliminary fusion result.

[0216] The fusion result is the second converged graph vector of the associated trajectory of the sample power equipment.

[0217] This second convergent graph vector comprehensively considers various factors in the sample power equipment association trajectory, as well as the relationship between the prior template vector and the sample power material knowledge vector.

[0218] For example, suppose this sample power equipment association trajectory represents the situation of a medium-sized thermal power station. The first association trend vector may reflect the demand trend of the power station for high-efficiency transformers, and its self-attention coefficient is high because the power station plans to improve power generation efficiency in the near future. The second association trend vector may be the demand trend for environmentally friendly cables, and its self-attention coefficient is low because environmental protection requirements are not the current priority. The third association trend vector may be the demand trend for low-cost capacitors, and the coefficient is high because operating cost control is more important. The fourth association trend vector may be the demand trend for new intelligent monitoring equipment, and the coefficient is low because it is currently in the preliminary consideration stage.

[0219] The second convergence graph vector generated through such fusion calculation can accurately reflect the comprehensive situation and key directions of the demand for various power equipment of this medium-sized thermal power station under the current circumstances.

[0220] The server performs the same processing on other sample power equipment-related trajectories to generate respective second converged graph vectors.

[0221] These second convergence graph vectors provide an important basis for subsequent analysis and decision-making, such as helping the server determine which power stations need to prioritize equipment updates, which power materials need to increase reserves, and how to optimize the equipment configuration and resource allocation of the entire power system.

[0222] In another specific example, if it is a sample power equipment associated trajectory of a hydroelectric power station, its associated trend vector may focus more on the maintenance and upgrade needs of turbines and special power materials related to hydroelectric power generation. Through processing and fusion with the prior template vector and the sample power material knowledge vector, the generated second converged graph vector can highlight the characteristics and needs of the hydroelectric power station.

[0223] This can provide accurate and valuable information support for the optimized management and decision-making of the power system.

[0224] In a possible implementation, step S131 includes:

[0225] Step S1311, extracting the knowledge vector of the target sample power material and the knowledge vector of each associated sample power station having an operational relationship with the target sample power material from the trajectory graph vector of the third unit trajectory, the third unit trajectory being one of the Y sample material associated trajectories, and the target sample power material being one of the sample power materials in the third unit trajectory;

[0226] Step S1312, performing self-attention processing on the knowledge vector of the target sample electric power material and the knowledge vector of each of the associated sample power stations, respectively, to generate a sample attention coefficient of each of the associated sample power stations in the third unit trajectory;

[0227] Step S1313 , performing fusion calculation on the knowledge vectors corresponding to the associated sample power stations according to the sample attention coefficient of each associated sample power station, and generating the knowledge logic graph feature of the third unit trajectory.

[0228] In this embodiment, one of the Y sample material-related trajectories is marked as a third unit trajectory.

[0229] In the trajectory graph vector of this third unit trajectory, the server first extracts the knowledge vector. Assume that this third unit trajectory involves a specific target sample power material, such as a certain type of smart meter, and there are multiple associated sample power plants that have an operational relationship with the target sample power material, such as power plant X, power plant Y, and power plant Z.

[0230] The server extracts the knowledge vector of the target sample power material (smart meter) from the trajectory graph vector of the third unit trajectory, which contains detailed information such as the technical specifications, production batches, and quality inspection data of the smart meter. At the same time, the server also extracts the knowledge vectors of power stations X, Y, and Z that have operational relationships with the smart meter. These knowledge vectors contain information such as the scale, power load, and power equipment configuration of each power station.

[0231] Next, the server performs self-attention processing on the knowledge vector of the target sample power material (smart meter) and the knowledge vector of each associated sample power station.

[0232] Taking power station X as an example, the server analyzes the matching degree and mutual influence between the characteristics of the smart meter and the operational needs of power station X through a series of complex calculations and comparisons. For example, it is found that power station X has high requirements for the accuracy and data transmission speed of the smart meter due to its high load of industrial electricity demand. Through this self-attention processing, the server generates a sample attention coefficient for power station X, which reflects the importance of the smart meter to power station X.

[0233] In the same way, the server also performs self-attention processing on power station Y and power station Z, and generates corresponding sample attention coefficients respectively. Then, the server performs fusion calculation on the knowledge vectors of the corresponding associated sample power stations according to the sample attention coefficients of each associated sample power station.

[0234] Assume that the sample attention coefficient of power station X is 0.7, the sample attention coefficient of power station Y is 0.6, and the sample attention coefficient of power station Z is 0.8. The server multiplies the knowledge vector of power station X by 0.7, the knowledge vector of power station Y by 0.6, and the knowledge vector of power station Z by 0.8. Then these products are added together to obtain a preliminary fusion result. This fusion result is the knowledge logic graph feature of the third unit trajectory.

[0235] For example, through the above processing, the server finds that due to the technical upgrade of smart meters, power station Z may need to plan in advance to replace the existing meters to meet more accurate power metering and monitoring needs. This discovery is reflected in the generated knowledge logic graph features, providing an important basis for subsequent power material management and power station operation decisions.

[0236] In another specific scenario, if the smart meter has remote meter reading and fault warning functions, and power station Y is located in a remote area with few operation and maintenance personnel, then the knowledge logic graph features generated by self-attention processing and fusion calculation will highlight the importance of these functions of the smart meter to power station Y, which may prompt power station Y to increase the purchase and installation of this model of smart meters.

[0237] Therefore, valuable knowledge logic graph features can be extracted from complex sample material association trajectories, providing strong support for material management and operation optimization of power systems.

[0238] In a possible implementation, step S110 includes: obtaining a demand-related knowledge graph, wherein the demand-related knowledge graph includes X sample power equipment-related trajectories and Y sample material-related trajectories, and the demand-related knowledge graph represents demand-related topology information between the sample power station and the sample power material.

[0239] In this embodiment, the server starts to execute the task of obtaining X sample power equipment association trajectories and Y sample material association trajectories. First, the server accesses the demand association knowledge graph stored in a large database.

[0240] This demand association knowledge graph is a complex and comprehensive structure that integrates a large amount of information to characterize the demand association topology information between sample power plants and sample power materials.

[0241] Assume that in this demand-related knowledge graph, a sample power station, such as Power Station A, has detailed records of its demand for various types of power equipment. Power Station A is a large hydroelectric power station that has a continuous demand for specific types of power equipment such as turbines, transformers, and transmission lines. The time series, quantity changes, equipment specification requirements, and other information of these demands are carefully organized in the knowledge graph, forming a track of power equipment related to Power Station A.

[0242] Let's look at another example power station, for example, Power Station B, which is an emerging solar power station. Due to its unique power generation method and operation mode, it has specific requirements for power materials such as photovoltaic panels, inverters, and energy storage batteries. These demand information also constitutes the power equipment association track of Power Station B in the knowledge graph.

[0243] For sample materials, a certain type of cable is used as an example. This type of cable has different application scenarios and demand frequencies in different power stations, such as power station C, power station D, and power station E. In the demand association knowledge graph, detailed information such as the purchase time, usage quantity, and installation location of this type of cable for each power station is recorded, thus forming the material association track of the sample material.

[0244] Through complex data retrieval and extraction algorithms, the server accurately obtains X sample power equipment association trajectories and Y sample material association trajectories from this huge demand association knowledge graph.

[0245] For example, the server may first select a group of related power stations from the knowledge graph based on specific screening conditions, such as the size, geographical location, and operating time of the power station. For each selected power station, its association information with various power equipment and materials is further extracted to form independent sample power equipment association trajectories.

[0246] Similarly, for power materials, the server can filter them according to the type, specification, manufacturer and other attributes of the materials, and then obtain the relevant power station demand information to form a sample material association track.

[0247] During the acquisition process, the server also needs to process various relationships and links in the knowledge graph. For example, some power equipment may be shared by multiple power stations at the same time, and the server needs to accurately track the flow of these equipment between different power stations and changes in demand to ensure that the acquired association track is complete and accurate.

[0248] Suppose during the acquisition process, the server discovers that a certain power plant has undergone a technological transformation, and the demand for a certain power material that was originally stable has suddenly increased significantly. Through in-depth mining of the knowledge graph, the server can trace the source of this change and find that it is because the new production process requires higher-performance power material support, thereby more comprehensively understanding the background and impact of this demand change.

[0249] For another example, for a new type of power equipment, the server analyzes the associated information in the knowledge graph and finds that it has been rapidly applied and promoted in some specific types of power stations, but has not yet been adopted in other power stations. This provides an important reference for the server to evaluate the market prospects and applicability of the new equipment.

[0250] That is, this embodiment extracts rich sample power equipment association trajectories and sample material association trajectories from the demand association knowledge graph, providing a solid data foundation for subsequent analysis, prediction, and decision-making.

[0251] Therefore, through interaction with the database and accurate analysis of the demand-related knowledge graph, a large number of valuable sample association trajectories can be effectively obtained, providing key information support for optimizing the operation and management of the power system.

[0252] In a possible implementation, the method further includes:

[0253] Step B110, obtaining X target power equipment associated trajectories and Y target material associated trajectories;

[0254] Step B120, extracting the trajectory map vector of each of the target power equipment associated trajectories and the trajectory map vector of each of the target materials associated trajectories;

[0255] Step B130, using the trajectory map vectors of the X target power equipment associated trajectories and the trajectory map vectors of the Y target material associated trajectories as model loading data of the target prediction neural network to generate a target power material knowledge vector and a target power station knowledge vector;

[0256] Step B140, constructing an association matrix based on the target power material knowledge vector and the target power station knowledge vector, wherein the elements of the association matrix represent the correlation measurement coefficient between the target power material knowledge vector and the target power station knowledge vector;

[0257] Step B150, using a graph search algorithm or a path planning algorithm to search for possible demand association links in the association matrix, and for each candidate demand association link searched, calculating a comprehensive relevance score of the candidate demand association link, the comprehensive relevance score being obtained by weighted summation based on the relevance metrics of each element in the candidate demand association link;

[0258] Step B160, screening and sorting the searched candidate demand-associated links according to the comprehensive relevance score, and selecting a preset number of candidate demand-associated links as reference demand-associated links based on the sorting result;

[0259] Step B170, obtaining a reference power material sequence, and classifying the reference power material sequence according to the type, specification, purpose and / or historical demand of the power material;

[0260] Step B180, matching the reference demand association link with the classified reference power material sequence, for each reference demand association link, traversing the reference power material sequence, searching for items related to the power materials and power stations involved in the reference demand association link, calculating the matching degree of the reference demand association with each matching item in the reference power material sequence, the matching degree being calculated based on the attribute similarity and location information of the item;

[0261] Step B190, selecting a final target power material demand result according to the matching degree of each matching item in the reference power material sequence associated with the reference demand.

[0262] In this embodiment, the server first accesses the big data platform of the power system, which integrates detailed data from multiple power stations and power material supplies across the country.

[0263] The server starts to obtain the target power equipment-related tracks and target material-related tracks through specific data screening conditions and query statements. For example, the server sets the conditions to obtain the power equipment-related tracks of 20 power stations located in the eastern region with installed capacity greater than a certain value. These tracks cover equipment procurement, maintenance and replacement records in the past six months. At the same time, 30 target material-related tracks are obtained, including the supply and use of materials such as wires, transformer components and insulators of different specifications in these power stations and surrounding areas.

[0264] Take one of the target power equipment association trajectories as an example. Power Plant E is a large nuclear power plant located on the eastern coast. In the past six months, it has successively purchased new nuclear reactor control equipment, upgraded the protection devices of some transmission lines, and regularly calibrated and repaired some key monitoring instruments. These detailed operations are recorded in its power equipment association trajectory, including specific operation time, equipment model, quantity, and related technical parameters.

[0265] For a target material association trajectory, for example, a certain specification of high temperature resistant wire F is widely used by power station E and several nearby small distributed power stations. Its supply time is scattered in different months, and the supply quantity varies according to the construction and maintenance plans of each power station. This information together constitutes the target material association trajectory of wire F.

[0266] Next, after receiving the acquired target power equipment-related trajectory data and target material-related trajectory data, the server starts to extract the trajectory map vector.

[0267] For the associated trajectory of power equipment of power plant E, the server first converts the geographical location (latitude and longitude) of the power plant into precise digital coordinates. For the purchased nuclear reactor control equipment, specific codes are assigned according to the equipment model and technical specifications. The purchase time of the equipment is converted into a precise timestamp, and the number of repairs and calibrations is recorded as an integer value. This information is arranged in a pre-set order and format to form a multi-dimensional trajectory map vector, which fully describes the dynamic situation of power plant E in terms of power equipment.

[0268] The server performs similar processing on the target material-related trajectory of the power line F. The specifications of the power line (such as cross-sectional area, material, and temperature tolerance) are encoded as a specific digital combination. The identification of each power station is converted into a unique digital code. The supply time is converted into a corresponding integer in months, and the supply quantity is directly recorded as a specific numerical value. Through such processing, the target material-related trajectory of the power line F is also converted into a trajectory graph vector that can be subsequently calculated and analyzed.

[0269] Next, the server will extract all the trajectory map vectors of the target power equipment-related trajectories and target material-related trajectories and input them into the trained target prediction neural network.

[0270] The neurons in the network begin to perform complex calculations and information processing on the large amount of input data. Taking one specific calculation process as an example, after the first layer of neurons in the network receives the trajectory map vectors of power equipment in multiple power stations, they extract some basic features through weight distribution and activation functions. Then, these features are passed to deeper network layers, and after multiple convolution, pooling and full connection operations, the deep patterns and rules hidden in the data are gradually excavated.

[0271] After multiple layers of calculation and transmission, the target power material knowledge vector and the target power station knowledge vector are finally generated. It is assumed that the generated target power material knowledge vector contains information such as the future demand trend of various power materials, potential supply bottlenecks, and the impact of technological development on material demand. The target power station knowledge vector reflects key knowledge such as the aging status of equipment in each power station, future expansion plans, and competitive advantages in the power market.

[0272] Next, the server starts to construct a correlation matrix to reveal the intrinsic relationship between the target power material knowledge vector and the target power plant knowledge vector.

[0273] Assuming that the target power material knowledge vector covers 8 major types of power materials, and the target power station knowledge vector includes 15 different power stations, the server determines the degree of correlation between each power material and each power station through complex calculations and analysis.

[0274] For example, for a certain high-performance transformer material, the server comprehensively considers factors such as the installed capacity of the power station, years of operation, recent equipment upgrade plans, and the growth trend of electricity demand in the region, and calculates its correlation measurement coefficient with power station 1 to be 0.75, and its correlation measurement coefficient with power station 2 to be 0.6, and so on, gradually filling the entire 8×15 association matrix.

[0275] Next, the server uses a graph search algorithm to explore potential demand-related links from the constructed association matrix.

[0276] For example, the server finds a link from a specific type of cable material to power station 7. Then, the server starts to calculate the comprehensive relevance score of this candidate demand association link. Assume that this link passes through 4 elements, and their relevance measurement coefficients are 0.8, 0.7, 0.9, and 0.6 respectively. The server assigns weights of 0.25, 0.3, 0.35, and 0.1 to these 4 coefficients respectively, and calculates the comprehensive relevance score through weighted summation: 0.8×0.25+0.7×0.3+0.9×0.35+0.6×0.1=0.795.

[0277] Next, the server carefully sorts all searched candidate demand-related links according to the calculated comprehensive relevance scores.

[0278] Assuming the preset number is 10, the server ranks the comprehensive relevance scores from high to low, and selects the top 10 candidate demand association links with the highest scores as demand association links with important reference value. For example, the top ones may be the link from advanced smart meter materials to power station 12, and the link from high-efficiency energy-saving transformer materials to power station 9.

[0279] Next, the server obtains a comprehensive reference power material sequence from the power system's material database, which may contain thousands of different power materials.

[0280] The server then classifies the reference power material sequence according to a series of criteria. For example, transformers, circuit breakers, wires, etc. are classified into different categories according to the type of power materials; transformers are further subdivided into subcategories of different capacities according to specifications; wires are divided into categories such as transmission and distribution based on usage; and materials are divided into high-demand, medium-demand, and low-demand categories based on historical demand data, such as procurement frequency and quantity in the past few years.

[0281] Taking transformers as an example, the server classifies all transformer materials into one category, and then further subdivides them according to capacity (such as 500KVA, 1000KVA, etc.). For wires, they are classified in detail according to their material (such as copper core, aluminum core), cross-sectional area, and withstand voltage level.

[0282] Next, the server accurately matches the previously screened reference demand association links with the classified reference power material sequence.

[0283] Taking a reference demand association link from an insulator material of a specific specification to power station 5 as an example, the server carefully traverses the entire reference power material sequence. When an insulator material of the same specification is found, its detailed attributes, such as the purity of the material, the specific value of the withstand voltage, etc., are further compared to determine the attribute similarity. At the same time, the distance between the geographical location of power station 5 and the material supply location is considered. Assuming that the attribute similarity is 85%, the location distance is close, and the location weight is assigned to 0.2, the calculated matching degree is 0.85×0.8+0.2×0.9=0.82.

[0284] For each matching item, the server repeats this calculation process to ensure accurate assessment of the matching degree between each reference demand association link and each item in the reference power material sequence.

[0285] Next, the server comprehensively considers the matching degree of all matching items, conducts in-depth analysis and decision-making to determine the final target power material demand results.

[0286] For example, the server sorts all the matching degrees and gives priority to the items with high matching degrees. Assume that after sorting, it is determined that the demand for insulators of a certain specification at power station 5 is the most urgent and has the highest matching degree, and power station 8 also has a high matching degree and urgent demand for a certain type of transformer material. The server sorts these results and determines them as the final target power material demand results.

[0287] These results will provide data support with practical guidance for the accurate procurement and rational allocation of power materials, as well as equipment renewal and maintenance plans for power plants, and help improve the operating efficiency and stability of the power system.

[0288] In this way, a series of complex tasks from acquiring data to finally determining the target power material demand results were completed, providing key decision-making basis for the optimized management and efficient operation of the power system.

[0289] Figure 2 The hardware structure of the deep learning system 100 provided in the embodiment of the present application for implementing the above-mentioned power material demand forecasting method based on deep learning is shown, as Figure 2 As shown, the deep learning system 100 may include a processor 110 , a machine-readable storage medium 120 , a bus 130 , and a communication unit 140 .

[0290] In an alternative embodiment, the deep learning system 100 may be a single server or a server group. The server group may be centralized or distributed (for example, the deep learning system 100 may be a distributed system). In an alternative embodiment, the deep learning system 100 may be local or remote. For example, the deep learning system 100 may access information and / or data stored in a machine-readable storage medium 120 via a network. For another example, the deep learning system 100 may be directly connected to the machine-readable storage medium 120 to access the stored information and / or data. In an alternative embodiment, the deep learning system 100 may be implemented on a cloud platform. By way of example only, the cloud platform may include a private cloud, a public cloud, an aggregated cloud, a community cloud, a distributed cloud, an internal cloud, a multi-layer cloud, or the like, or any integration thereof.

[0291] The machine-readable storage medium 120 may store data and / or instructions. In an alternative embodiment, the machine-readable storage medium 120 may store data obtained from an external terminal. In an alternative embodiment, the machine-readable storage medium 120 may store data and / or instructions used by the deep learning system 100 to execute or use to complete the exemplary method described in this application. In an alternative embodiment, the machine-readable storage medium 120 may include a mass storage, a removable storage, a volatile read-write memory, a read-only memory, etc. or any integration thereof. Exemplary mass storage may include a magnetic disk, an optical disk, a solid-state disk, etc. Exemplary removable storage may include a flash drive, a floppy disk, an optical disk, a memory card, a compressed disk, a magnetic tape, etc.

[0292] During the specific implementation process, multiple processors 110 execute computer executable instructions stored in the machine-readable storage medium 120, so that the processor 110 can execute the power material demand forecasting method based on deep learning in the above method embodiment. The processor 110, the machine-readable storage medium 120 and the communication unit 140 are connected through the bus 130, and the processor 110 can be used to describe the sending and receiving actions of the communication unit 140.

[0293] The specific implementation process of the processor 110 can be found in the various method embodiments executed by the above-mentioned deep learning system 100. The implementation principles and technical effects are similar and will not be repeated in this embodiment.

[0294] In addition, an embodiment of the present application also provides a readable storage medium, in which computer executable instructions are pre-installed. When the processor executes the computer executable instructions, the above-mentioned method for predicting power material demand based on deep learning is implemented.

[0295] Similarly, it should be noted that in order to simplify the description disclosed in this application, thereby helping to understand one or more embodiments of the invention, in the foregoing description of the embodiments of the present application, multiple features are sometimes combined into one embodiment, drawings, or descriptions thereof. Similarly, it should be noted that in order to simplify the description disclosed in this application, thereby helping to understand one or more embodiments of the invention, in the foregoing description of the embodiments of the present application, multiple features are sometimes combined into one embodiment, drawings, or descriptions thereof.

Claims

1. A method for predicting power material demand based on deep learning, characterized in that: The method comprises: obtaining X sample power equipment association trajectories and Y sample material association trajectories, each of the sample power equipment association trajectories representing a demand association knowledge point between the sample power station and the sample power material when the sample power station is used as the demand guiding node, and each of the sample material association trajectories representing a demand association knowledge point between the sample power material and the sample power station when the sample power material is used as the demand guiding node; the sample power equipment association trajectory refers to a data point set recording the demand association situation between the sample power station and various power equipment with the sample power station as the demand guiding node in the power system, and the data points include the time, quantity, equipment type and specification of the demand; the sample material association trajectory is based on The sample power material is a demand-guiding node, which records a set of data points of the demand association between the sample power material and different power stations. The data points include the supply time, quantity, specification, and receiving power station of the sample power material, reflecting the actual demand and use of the sample power material by different power stations; extracting the trajectory map vector of each of the sample power equipment associated trajectories and the trajectory map vector of each of the sample material associated trajectories; using the trajectory map vectors of the Y sample material associated trajectories as the model loading data of the first prediction unit in the basic prediction neural network to generate a sample power material knowledge vector, and using one of the sample power material knowledge vector or the first derivative transfer vector, X sample power equipment associated trajectories as the model loading data of the first prediction unit in the basic prediction neural network to generate a sample power material knowledge vector, and The first derived transfer vector represents the feature vector of the second prediction unit when fitting the prediction result of the first prediction unit; the feature distance between the sample power material knowledge vector and the sample power plant knowledge vector is calculated to obtain the target training cost parameter, and the basic prediction neural network is subjected to parameter learning according to the target training cost parameter to generate the target prediction neural network, and the target prediction neural network is used to predict and output the trajectory map vectors of the X target power equipment associated trajectories and the trajectory map vectors of the Y target material associated trajectories, and generate the target power material knowledge vector and a target power station knowledge vector, the target power material knowledge vector and the target power station knowledge vector are used to determine the target power material demand result; the first prediction unit includes a first fusion subunit and a first focus allocation subunit; the trajectory graph vectors of the Y sample material associated trajectories are used as model loading data of the first prediction unit in the basic prediction neural network to generate a sample power material knowledge vector, including: loading the trajectory graph vector of each sample material associated trajectory into the first fusion subunit respectively, generating a knowledge logic graph feature of each sample material associated trajectory, the knowledge logic graph feature of each sample material associated trajectory representing the logical feature information associated with the material demand in the sample material associated trajectory;The knowledge logic graph features of the Y sample material associated trajectories are used as the model loading data of the first focus allocation subunit to generate the focus attention coefficient of each sample material associated trajectory; the knowledge logic graph features of the Y sample material associated trajectories and the focus attention coefficient are fused and calculated according to the first focus allocation subunit to generate a sample power material knowledge vector; before the trajectory graph vector of each sample material associated trajectory is loaded into the first fusion subunit respectively to generate the knowledge logic graph features of each sample material associated trajectory, the method also includes: obtaining a second derivative transfer vector, the second derivative transfer vector represents the prediction of the first prediction unit fitting the prediction of the second prediction unit The feature vector of the result; loading the trajectory graph vector of each sample material associated trajectory into the first fusion subunit respectively, generating the knowledge logic graph feature of each sample material associated trajectory, including: converging the second derivative transfer vector with the trajectory graph vector of each sample material associated trajectory respectively, generating the converged graph vector of each sample material associated trajectory; loading the converged graph vector of each sample material associated trajectory into the first fusion subunit respectively, generating the knowledge logic graph feature of each sample material associated trajectory; wherein, according to the first focus allocation subunit, the knowledge logic graph features of Y sample material associated trajectories are fused and calculated with the focus attention coefficient, Generate a sample power material knowledge vector, including: according to the first focus allocation subunit, fuse and calculate the knowledge logic graph features of Y sample material associated trajectories with the focus attention coefficient to generate a target convergence graph vector; integrate the target convergence graph vector with the second derived transfer vector to generate a sample power material knowledge vector; the second prediction unit includes a second fusion subunit and a second focus allocation subunit; use the sample power material knowledge vector or one of the first derived transfer vectors and the trajectory graph vector of X sample power equipment associated trajectories as model loading data of the second prediction unit in the basic prediction neural network to obtain a sample power station knowledge vector, including: the first derived transfer vector The handover vector is respectively integrated with the trajectory graph vector of each of the sample power equipment associated trajectories to generate a first converged graph vector of each of the sample power equipment associated trajectories; the first converged graph vector of each of the sample power equipment associated trajectories is respectively loaded into the second fusion subunit to generate a knowledge logic graph feature of each of the sample power equipment associated trajectories, and the knowledge logic graph feature of each of the sample power equipment associated trajectories represents the logical feature information of the power equipment associated in the corresponding sample power equipment associated trajectories; the knowledge logic graph feature of each of the sample power equipment associated trajectories is used as the model loading data of the second focus allocation subunit to generate a focus attention coefficient of each of the sample power equipment associated trajectories;According to the second focus allocation subunit, the knowledge logic graph features of the X sample power equipment associated trajectories and the focus attention coefficient are fused and calculated to generate a first focused fusion knowledge vector; the first focused fusion knowledge vector is integrated with the first derivative transfer vector to generate a sample power plant knowledge vector; wherein the first convergence graph vector of each sample power equipment associated trajectory is loaded into the second fusion subunit respectively to generate the knowledge logic graph feature of each sample power equipment associated trajectory, including: extracting the knowledge vector of the target sample power plant and the knowledge vector of each associated sample power plant that has an operational relationship with the target sample power plant from the first convergence graph vector of the first unit trajectory, the first unit trajectory is one of the X sample power equipment associated trajectories, and the target sample power plant is one of the sample power plants in the first unit trajectory; performing self-attention processing on the knowledge vector of the target sample power plant and the knowledge vector of each associated sample power plant respectively to generate the sample attention coefficient of each associated sample power plant in the first unit trajectory; performing fusion calculation on the knowledge vector corresponding to the associated sample power plant according to the sample attention coefficient of each associated sample power plant to generate the knowledge logic graph feature of the first unit trajectory. ; 2. The method for predicting power material demand based on deep learning according to claim 1, characterized in that: The method also includes: calculating a first loss function value between the sample power plant knowledge vector and the second derived transfer vector; adjusting the first derived transfer vector and the second derived transfer vector based on the first loss function value; and, calculating a second loss function value between the sample power material knowledge vector and the first derived transfer vector; adjusting the second derived transfer vector and the first derived transfer vector based on the second loss function value.

3. The method for predicting power material demand based on deep learning according to claim 1, characterized in that: The second prediction unit includes Z graph autoencoders, a gated selection subunit and a multilayer perceptron; the sample power material knowledge vector or one of the first derived transfer vectors and the trajectory graph vectors of the X sample power equipment associated trajectories are used as model loading data of the second prediction unit in the basic prediction neural network to obtain a sample power station knowledge vector, including: loading the trajectory graph vector of each sample power equipment associated trajectory and Z prior template vectors into the Z graph autoencoders, generating Z associated trend vectors of each sample power equipment associated trajectory, each graph autoencoder generates one associated trend vector, and each prior template vector reflects the corresponding graph autoencoder The prior attention weight of the sample power material is determined by the device; the Z associated trend vectors of each of the sample power equipment associated trajectories, the Z prior template vectors and the sample power material knowledge vector are loaded into the gated selection subunit to generate a second converged graph vector of each of the sample power equipment associated trajectories; the second converged graph vectors of the X sample power equipment associated trajectories are loaded into the multi-layer perceptron to generate a sample power station knowledge vector; wherein the trajectory graph vector of each of the sample power equipment associated trajectories and the Z prior template vectors are loaded into the Z graph autoencoders to generate Z associated trend vectors of each of the sample power equipment associated trajectories, including: from the second unit trajectory In the trajectory graph vector, the knowledge vector of the target sample power station and the knowledge vector of each associated sample power station that has an operational relationship with the target sample power material are extracted, the second unit trajectory is one of the X sample power equipment associated trajectories, and the target sample power station is one of the sample power materials in the second unit trajectory; the knowledge vector of the target sample power station is respectively converged with the Z prior template vectors to generate Z first converged vectors; the Z first converged vectors are respectively processed with the knowledge vector of each associated sample power station by self-attention to generate a sample attention coefficient of each associated sample power station in the second unit trajectory; according to each associated The sample attention coefficient of the sample power station performs a fusion calculation on the knowledge vector corresponding to the associated sample power station to generate Z associated trend vectors of the second unit trajectory; wherein the Z associated trend vectors of each of the sample power equipment associated trajectories, the Z prior template vectors and the sample power material knowledge vector are loaded into the gated selection subunit to generate a second convergence graph vector of each of the sample power equipment associated trajectories, including: for each of the sample power equipment associated trajectories, the Z prior template vectors are respectively self-attention processed with the sample power material knowledge vector to generate a self-attention coefficient of each of the associated trend vectors in each of the sample power equipment associated trajectories;For each of the sample power equipment associated trajectories, the corresponding Z associated trend vectors are fused and calculated according to the self-attention coefficients of the Z associated trend vectors to generate a second converged graph vector of each of the sample power equipment associated trajectories. ; 4. The method for predicting power material demand based on deep learning according to claim 1, characterized in that: The trajectory graph vectors of each of the sample material associated trajectories are loaded into the first fusion subunit respectively, and the knowledge logic graph features of each of the sample material associated trajectories are generated, including: extracting the knowledge vector of the target sample power material and the knowledge vector of each associated sample power station that has an operational relationship with the target sample power material from the trajectory graph vector of the third unit trajectory, the third unit trajectory is one of the Y sample material associated trajectories, and the target sample power material is one of the sample power materials in the third unit trajectory; performing self-attention processing on the knowledge vector of the target sample power material and the knowledge vector of each of the associated sample power stations respectively, to generate a sample attention coefficient of each of the associated sample power stations in the third unit trajectory; performing fusion calculation on the knowledge vector corresponding to the associated sample power station according to the sample attention coefficient of each of the associated sample power stations, to generate the knowledge logic graph features of the third unit trajectory.

5. The method for predicting power material demand based on deep learning according to claim 1, characterized in that: Obtaining X sample power equipment association trajectories and Y sample material association trajectories includes: obtaining a demand association knowledge graph, the demand association knowledge graph includes X sample power equipment association trajectories and Y sample material association trajectories, and the demand association knowledge graph represents demand association topology information between the sample power station and the sample power material.

6. The method for predicting power material demand based on deep learning according to claim 1, characterized in that: The method further includes: obtaining X target power equipment associated trajectories and Y target material associated trajectories; extracting a trajectory map vector of each target power equipment associated trajectory and a trajectory map vector of each target material associated trajectory; using the trajectory map vectors of the X target power equipment associated trajectories and the trajectory map vectors of the Y target material associated trajectories as model loading data of the target prediction neural network to generate a target power material knowledge vector and a target power station knowledge vector; constructing an association matrix based on the target power material knowledge vector and the target power station knowledge vector, wherein the elements of the association matrix represent the correlation measurement coefficient between the target power material knowledge vector and the target power station knowledge vector; searching for possible demand association links in the association matrix using a graph search algorithm or a path planning algorithm, and for each candidate demand association link searched, calculating a comprehensive correlation score of the candidate demand association link, wherein the comprehensive correlation score is based on the correlation coefficients of each candidate demand association link in the candidate demand association link. The weighted summation of the correlation metrics of the elements is obtained; according to the comprehensive correlation score, the searched candidate demand association links are screened and sorted, and a preset number of candidate demand association links are selected as reference demand association links based on the sorting results; a reference electric power material sequence is obtained, and the reference electric power material sequence is classified according to the type, specification, purpose and / or historical demand of the electric power material; the reference demand association link is matched with the classified reference electric power material sequence, and for each reference demand association link, the reference electric power material sequence is traversed to find the projects related to the electric power materials and power stations involved in the reference demand association link, and the matching degree of each matching project in the reference electric power material sequence of the reference demand association is calculated, and the matching degree is calculated based on the attribute similarity and location information of the project; according to the matching degree of each matching project in the reference electric power material sequence of the reference demand association, the final target electric power material demand result is selected.

7. A deep learning system, characterized in that: The deep learning system includes a processor and a machine-readable storage medium, wherein the machine-readable storage medium stores machine-executable instructions, and the machine-executable instructions are loaded and executed by the processor to implement the deep learning-based electricity material demand forecasting method described in any one of claims 1-6.

Images