Major network load multi-source data fusion prediction method based on deep learning

Abstract

The invention discloses a main network load multi-source data fusion prediction method based on deep learning. The method comprises the following steps of S1, performing alignment, complementation and normalization preprocessing on multi-source data; s2, constructing a modal embedding structure, and generating modal embedding representation with unified dimensions; s3, extracting local features by using multi-scale convolution; s4, introducing a time sequence convolutional network of a dynamic routing control mechanism, and generating and updating a modal routing matrix; s5, scheduling a channel structure to construct a candidate path; s6, coding the candidate structure to construct a structure optimization space; s7, adopting an improved tornado optimization algorithm to perform structure search and evolution; and S8, completing load prediction based on the optimal structure and outputting a result. According to the method, the modal dynamic modeling and the structure-adjustable convolution path are combined, the efficient optimization strategy is matched, high-precision and explainable main network load prediction is realized, and the method has good engineering adaptability and popularization value.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power system load forecasting and artificial intelligence, and particularly relates to a main grid load multi-source data fusion forecasting method based on deep learning. BACKGROUND

[0002] With the continuous improvement of the intelligentization of power grids and the fine degree of dispatching, as an important basis for realizing the safe operation and optimal control of power grids, the modeling accuracy and real-time response capability of main grid load forecasting have attracted widespread attention. In recent years, deep learning technology has been gradually introduced into this field, and by fusing multi-source data such as historical load, weather changes and operating environment, the modeling capability of the model for complex time series changes has been improved. At the same time, multi-modal feature representation, convolution structure enhancement and adaptive optimization strategy have also become the research focus of improving load forecasting effect. The structure represented by the time series convolution network (TCN) has shown excellent feature extraction capability in processing power load sequences, and to some extent, has made up for the problems of gradient dissipation and long dependence modeling difficulty existing in traditional RNN models.

[0003] However, in actual application, the existing main grid load forecasting methods generally have the following problems: the multi-source heterogeneous data are significantly different in time dimension, modal type and numerical scale, and the conventional preprocessing operations are difficult to achieve high-fidelity time alignment, abnormality elimination and normalization mapping, which affects the feature learning quality of the subsequent model; the modal embedding process lacks structure distinction, and it is difficult to accurately reflect the expression difference between numerical and categorical data; the channel structure of the time series convolution network is fixed, and lacks a dynamic connection regulation mechanism based on modal response, which limits the modeling capability of the change of modal contribution in the load fluctuation process; in addition, the generation and optimization process of candidate structures depends on manual experience or static search strategy, and lacks an efficient and global structure search mechanism, and traditional optimization algorithms such as genetic algorithm or particle swarm optimization have problems such as being easy to fall into local optimum, being insensitive to disturbance, and converging slowly when facing high-dimensional structure space. These technical bottlenecks significantly restrict the adaptability and stability of the main grid load forecasting model in complex scenarios.

[0004] Therefore, how to provide a main grid load multi-source data fusion forecasting method based on deep learning is a problem that those skilled in the art need to solve. SUMMARY

[0005] One purpose of the present application is to propose a deep learning-based main network load multi-source data fusion prediction method, the present application adopts a unified dimension mode embedding, convolution feature extraction, dynamic routing control mechanism time sequence convolution network structure, combines candidate structure generation and structure optimization space construction, and introduces an improved tornado optimization algorithm to search and update the fusion path, realizes the structure adaptive modeling and high-precision prediction of the main network load in the multi-modal input scene, and has the advantages of strong fusion expression ability, dynamic structure adjustment and efficient optimization mechanism.

[0006] According to the deep learning-based main network load multi-source data fusion prediction method of the embodiment of the present application, the following steps are included:

[0007] S1, collect multi-source input data, and perform time alignment, missing filling and normalization processing;

[0008] S2, perform embedding mapping on each modal data to generate a unified dimension mode embedding representation, and complete feature alignment in time sequence;

[0009] S3, input the mode embedding representation into the convolutional encoder, extract the local features of each mode through multi-scale convolution, and form a time sequence feature sequence;

[0010] S4, input the time sequence feature sequence into the time sequence convolution network with dynamic routing control mechanism, and generate a mode routing matrix according to the response mode of each modal channel in the time sequence feature sequence during the convolution modeling process, dynamically control the connection state and propagation path of each modal channel;

[0011] S5, schedule the convolution channel structure based on the mode routing matrix, construct the candidate prediction path, and form the corresponding candidate structure combination;

[0012] S6, construct a structure optimization space based on the candidate structure, the structure optimization space includes a mode routing matrix, a channel connection mode and a fusion parameter configuration;

[0013] S7, search and update in the structure optimization space using an improved tornado optimization algorithm, optimize the channel connection mode and fusion parameter configuration in the candidate structure combination, and determine the optimal prediction path structure;

[0014] S8, use the optimal prediction path structure for the load prediction task, and output the main network load prediction result of the target time period.

[0015] Optionally, the S1 specifically includes:

[0016] S11, for samples with time missing, the weighted linear interpolation algorithm is used to complete the data based on the data of the adjacent time points, wherein the interpolation result is calculated by weighted average of the front and rear effective values according to the time distance;

[0017] S12, for the abnormal value with mutation, a sliding median combined with range limit is used to remove, wherein the median in the sliding window of specified length is calculated as the reference, if the current sample deviates from the median more than the set range threshold, the sample is determined as abnormal and replaced by the window median;

[0018] S13, the maximum and minimum value normalization method is used to map the modal data to a unified numerical interval.

[0019] Optionally, the S2 specifically comprises:

[0020] S21, data type identification is performed on the modal data, the continuous variable therein is marked as numerical type modal data, and the discrete variable is marked as category type modal data, and is respectively assigned to an embedding submodule;

[0021] S22, for the numerical type modal data, an embedding structure comprising two layers of full connection network is constructed, the first layer uses ReLU activation function, and the second layer maps the original data to a preset unified embedding dimension space through linear transformation to generate numerical type embedding representation;

[0022] S23, for the category type modal data, a lookup table encoding combined with position encoding embedding method is used, the category label is first mapped to an index vector by using the lookup table, then the index vector is spliced with the position encoding vector corresponding to the time step, and a specified dimension category type embedding representation is output through a linear layer;

[0023] S24, the numerical type embedding representation and the category type embedding representation are frame by frame aligned according to the time index, the embedding sequence with time difference is resampled by using cubic spline interpolation, and each modal embedding representation is spliced and fused at each time step to generate a modal embedding representation of unified dimension.

[0024] Optionally, the S3 specifically comprises:

[0025] S31, a plurality of parallel convolution channels of the convolutional encoder are initialized, each convolution channel is configured with a time convolution kernel of different size, and the channel output dimension is set to a unified length;

[0026] S32, the modal embedding representation is input into each convolution channel, one-dimensional convolution operation is performed in each channel to extract sliding feature response at a corresponding scale;

[0027] S33, a batch normalization layer is sequentially connected after the output of each convolution channel, and a ReLU activation function is set to perform nonlinear mapping processing on the convolution result;

[0028] S34, the processed feature vectors in each convolution channel are spliced according to the channel dimension to form a time sequence feature sequence in a unified format.

[0029] Optionally, the S4 specifically comprises:

[0030] S41, initializing a time sequence convolution network with a multi-modal channel structure, inputting the time sequence feature sequence into the corresponding modal channel, each channel corresponding to an embedded modal;

[0031] S42, performing one-dimensional convolution operation in each modal channel, extracting local activation response features, and calculating the activation mean of the channel at the current time step;

[0032] S43, the activation mean of each modal channel forms an initial modal attention vector, and a Softmax operation is performed to generate a normalized modal attention coefficient;

[0033] S44, according to the normalized modal attention coefficient and the corresponding modal gating parameter, an element-by-element multiplication operation is performed to form a modal connection strength matrix, and a threshold pruning operation is performed to retain high weight items to generate a modal routing matrix;

[0034] S45, according to the modal routing matrix, the connection relationship between each modal channel in the time sequence convolution network is dynamically configured, the channel path with connection strength lower than the threshold is closed, and the high correlation connection is retained, realizing the structure update of the channel activation state in the current time step;

[0035] S46, performing time sequence convolution propagation operation on the updated connection structure to complete the dynamic modeling process of the current time step.

[0036] Optionally, the S5 specifically comprises:

[0037] S51, receiving the modal routing matrix and performing a multi-threshold segmentation operation to obtain a plurality of levels of connection strength masks to form a connection mask set;

[0038] S52, according to the connection mask set, combining the preset modal channel connection rule, enumerating a plurality of candidate prediction path structures, each scheme describing the on-off combination state of a group of modal channels;

[0039] S53, for each candidate prediction path structure, reconstructing its corresponding time sequence convolution propagation path structure, and establishing a structure index table to record its connection topology, activation channel and propagation direction information;

[0040] S54, based on each candidate prediction path structure, simulating the one-dimensional convolution propagation process on the original time sequence feature sequence to obtain the corresponding channel output state;

[0041] S55, calculate evaluation indexes of the output state of each candidate prediction path structure, including performance factors such as feature expression completeness, path response strength and structure sparsity, and generate a structure score vector;

[0042] S56, combine each candidate prediction path structure with its corresponding connection scheme, modal routing matrix parameters and structure score vector to form a candidate structure combination.

[0043] Optionally, the S6 specifically includes:

[0044] S61, extract the modal routing matrix, channel connection mode and fusion parameter configuration from the candidate structure combination;

[0045] S62, perform row vector expansion and normalization processing on the modal routing matrix to generate a one-dimensional sparse connection vector representing the connection strength distribution between each modality;

[0046] S63, construct a connection index vector for the channel connection mode to record the connection number and timing relationship between each pair of active channels, and obtain a channel structure code;

[0047] S64, extract the convolution kernel size, connection weight ratio and fusion operation type number from the fusion parameter configuration to generate a fusion parameter code vector;

[0048] S65, concatenate the one-dimensional sparse connection vector, channel structure code and fusion parameter code in a fixed order to generate a structure vector representation;

[0049] S66, combine all structure vector representations to form a structure optimization space.

[0050] Optionally, the S7 specifically includes:

[0051] S71, initialize the improved tornado optimization algorithm in the structure optimization space, take the structure vector as the population individual, set the maximum number of iterations, population size, upper and lower limits of inertia weight, structure guide coefficient, disturbance amplitude factor and mutation probability, and evaluate the initial individual with the fitness function, and select the individual with the optimal fitness as the initial center body;

[0052] S72, in each iteration, guide and update each wind body individual according to the structure difference between the center body and the current individual and the structure disturbance direction, and adjust the structure vector using the following formula:

[0053]

[0054] wherein, represents the structure vector of the current individual, represents the center body of the current round, r ％is the structure guiding coefficient, γ is the disturbance factor, is the structure disturbance gradient, used to enhance the adjustment ability of important structure dimensions;

[0055] S73, dynamically calculate the inertia weight according to the current iteration number, and scale the update result according to the inertia coefficient, so that the search step decreases with the iteration, and the inertia coefficient calculation formula is:

[0056]

[0057] w (t) represents the inertia coefficient at the tth iteration, represents the search step control coefficient of the current individual update, w ma* represents the initial maximum inertia coefficient, controls the search range in the initial stage of the algorithm, w min represents the minimum inertia coefficient, used to limit the search step in the late convergence, T is the maximum iteration number, represents the total number of rounds to be executed in the whole optimization process, t is the current iteration round, and the range is 0≤t≤T;

[0058] S74, perform structure disturbance operation on the individual with a set mutation probability, randomly select another population individual and introduce the structure difference into the current individual, and the update formula is:

[0059]

[0060] wherein, is the structure of the randomly selected individual, r2 is the mutation amplitude factor, and this mechanism is used to enhance the diversity of the population and the ability to jump out of the local optimum;

[0061] S75, recalculate the fitness of all updated individuals, and the fitness function is defined as:

[0062] F(x) = α·Strength(x) - β·Sparsity(x) - δ·Error(x);

[0063] wherein, Strength(x) represents the response strength of the structure path, defined as the time average of the weighted sum of the output activation value and the weight of each convolution channel; Sparsity(x) represents the sparsity of the structure connection, calculated as the proportion of the number of unconnected modes in the routing matrix to the total number of connections; Error(x) represents the prediction residual, which is the mean square error between the prediction result and the true load value; α, β, δ are the weighting coefficients of each part;

[0064] S76, update the central body structure according to the fitness result, if the maximum iteration number is reached or the optimal structure does not change in continuous multiple iterations, the optimization process is terminated, and the structure vector corresponding to the individual with the optimal fitness is output as the optimal prediction path structure.

[0065] Optionally, S8 specifically includes:

[0066] S81, configure the optimal prediction path structure as the structure of the current prediction model, which contains a modal routing matrix, a channel connection mode and a fusion parameter;

[0067] S82, input the time sequence feature sequence into the prediction model, and perform a forward propagation operation according to the channel connection relationship and the fusion mode defined by the optimal prediction path structure, to generate a prediction output sequence corresponding to a target time period;

[0068] S83, perform dimension reduction and inverse normalization processing on the prediction output sequence to restore the main grid load prediction result consistent with the original load data dimension, and output to the target interface according to the set time granularity.

[0069] The beneficial effects of the present application are:

[0070] (1) By constructing an embedding mapping mechanism that distinguishes between modal types, numerical data is continuously mapped using a two-layer fully connected structure, and a lookup table encoding and time position encoding fusion strategy is introduced for categorical data. After embedding, time alignment is completed using cubic spline interpolation, which significantly improves the fusion accuracy of original multi-source data at different time granularities and modal formats, effectively solving the problems of difficulty in aligning multi-modal data and insufficient expression ability in traditional load prediction methods.

[0071] (2) In the process of time series convolution modeling, a modal routing matrix generated based on modal attention and gating mechanism is introduced to dynamically control the connection state and propagation path of different modal channels, and combined with modal activation response to perform channel structure update, so that the convolution channel has time-varying adaptive ability, effectively overcoming the limitations of fixed model structure and inability to respond to changes in input modal importance, improving the structural flexibility and modeling accuracy of the model under load fluctuation conditions.

[0072] (3) By constructing a structure optimization space containing a modal routing matrix, a channel connection mode and a fusion parameter configuration, and introducing an improved tornado optimization algorithm with structure guidance, inertia wheel adjustment and random disturbance mechanism, global efficient search and update are realized in the candidate structure combination, which makes up for the weaknesses of traditional methods such as genetic algorithm and particle swarm in high-dimensional structure space, such as weak disturbance amplitude, slow convergence speed and easy to fall into local optimum, significantly improving the quality and optimization convergence efficiency of the prediction structure combination. BRIEF DESCRIPTION OF DRAWINGS

[0073] The accompanying drawings are included to provide a further understanding of the application, and constitute a part of the specification, together with the embodiments of the application, to explain the application, and do not constitute a limitation on the application. In the drawings:

[0074] Figure 1 A flowchart of a deep learning-based multi-source data fusion prediction method for main grid load is provided for the present application. DETAILED DESCRIPTION

[0075] The present application will now be further described in detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams and only schematically illustrate the basic structure of the present application, and thus only show the configurations related to the present application.

[0076] REFERENCE Figure 1 A deep learning-based multi-source data fusion prediction method for main grid load includes the following steps:

[0077] S1, collect multi-source input data and perform time alignment, missing data filling and normalization processing;

[0078] S2, perform embedding mapping on each modal data to generate modal embedding representation of uniform dimension, and complete feature alignment in time sequence;

[0079] S3, input the modal embedding representation into the convolutional encoder, extract the local features of each modal through multi-scale convolution, and form the time sequence feature sequence;

[0080] S4, input the time sequence feature sequence into the time sequence convolutional network with dynamic routing control mechanism, and generate the modal routing matrix according to the response mode of each modal channel in the time sequence feature sequence during the convolution modeling process, dynamically control the connection state and propagation path of each modal channel;

[0081] S5, schedule the convolution channel structure based on the modal routing matrix, construct the candidate prediction path, and form the corresponding candidate structure combination;

[0082] S6, construct a structure optimization space based on the candidate structure, which includes the modal routing matrix, the channel connection mode and the fusion parameter configuration;

[0083] S7, search and update in the structure optimization space using the improved tornado optimization algorithm, optimize the channel connection mode and the fusion parameter configuration in the candidate structure combination, and determine the optimal prediction path structure;

[0084] S8, use the optimal prediction path structure for the load prediction task, and output the main grid load prediction result of the target time period.

[0085] In the present embodiment, the S1 specifically includes:

[0086] S11, for samples with time missing, the weighted linear interpolation algorithm is used to complete the data based on the data of the adjacent time points, wherein the interpolation result is calculated by weighted average of the front and rear effective values according to the time distance;

[0087] S12, for the abnormal value with mutation, a sliding median combined with range limit is used to remove, wherein the median in the specified length of the sliding window is calculated as the reference, if the current sample deviates from the median more than the set range threshold, the sample is determined as abnormal and replaced with the window median;

[0088] S13, the maximum and minimum value normalization method is used to map the modal data to a unified numerical interval.

[0089] The embodiment fills in the missing data by interpolation, detects and replaces the abnormal value, and normalizes each modal data, thereby improving the consistency and stability of the multi-source data, and providing a high-quality input basis for subsequent embedding mapping and deep modeling.

[0090] In the embodiment, the S2 specifically comprises:

[0091] S21, data type identification is performed on each modal data, the continuous variable therein is marked as numerical type modal data, the discrete variable is marked as category type modal data, and is respectively assigned to an embedding submodule;

[0092] S22, for the numerical type modal data, an embedding structure comprising two layers of fully connected networks is constructed, the first layer uses a ReLU activation function, and the second layer maps the original data to a preset unified embedding dimension space through linear transformation to generate a numerical type embedding representation;

[0093] S23, for the category type modal data, an embedding method combining lookup table encoding and position encoding is used, the category label is first mapped to an index vector using a lookup table, then the index vector is spliced with a position encoding vector corresponding to the time step, and a specified dimension category type embedding representation is output through a linear layer;

[0094] S24, the numerical type embedding representation and the category type embedding representation are frame-aligned according to the time index, the embedding sequence with time difference is resampled by using cubic spline interpolation, and each modal embedding representation is spliced and fused at each time step to generate a modal embedding representation of a unified dimension.

[0095] The embodiment generates a modal embedding representation of a unified dimension by designing an embedding structure for numerical type and category type modal data respectively, and fusing position encoding and interpolation alignment strategy, thereby not only enhancing the expression ability of multi-modal features, but also improving the alignment accuracy between different time sequence modalities and the model perception effect.

[0096] In the embodiment, the S3 specifically comprises:

[0097] S31, a plurality of parallel convolution channels of the convolutional encoder are initialized, each convolution channel is configured with a time convolution kernel of different size, and the channel output dimension is set to a unified length;

[0098] S32, input the modal embedding representation into each convolution channel respectively, perform one-dimensional convolution operation in each channel, and extract sliding feature response at a corresponding scale;

[0099] S33, sequentially connect a batch normalization layer after the output of each convolution channel, and set a ReLU activation function to perform nonlinear mapping processing on the convolution result;

[0100] S34, perform a splicing operation on the processed feature vectors in each convolution channel according to the channel dimension to form a time sequence feature sequence in a unified format.

[0101] In the embodiment, the numerical modal is mapped using a full connection network, the categorical modal is combined with a lookup table encoding and a position encoding, and an interpolation alignment is used to generate a modal embedding representation in a unified dimension, which effectively improves the expression consistency and time sequence alignment accuracy of heterogeneous data, and provides more accurate and efficient feature input for model modeling.

[0102] In the embodiment, the S4 specifically includes:

[0103] S41, initialize a time sequence convolution network with a multi-modal channel structure, input a time sequence feature sequence into a corresponding modal channel, and each channel corresponds to an embedded modal one by one;

[0104] S42, perform one-dimensional convolution operation in each modal channel, extract local activation response features, and calculate the activation mean of the channel at the current time in time steps;

[0105] S43, form an initial modal attention vector by combining the activation means of each modal channel, and perform a Softmax operation to generate a normalized modal attention coefficient;

[0106] S44, perform element-wise multiplication operation according to the normalized modal attention coefficient and the gating parameter of the corresponding modal to form a modal connection strength matrix, and reserve high weight items through threshold pruning operation to generate a modal routing matrix;

[0107] S45, dynamically configure the connection relationship between each modal channel in the time sequence convolution network according to the modal routing matrix, close the channel path with a connection strength lower than the threshold, and reserve high correlation connection to realize the structure update of the channel activation state in the current time step;

[0108] S46, perform time sequence convolution propagation operation on the updated connection structure to complete the dynamic modeling process of the current time step.

[0109] The embodiment forms a modal routing matrix, dynamically adjusts the connection structure between each modal channel, updates the connection state at each time step, and reconstructs the propagation path, effectively solving the problem of fixed and non-adaptive input feature changes of the traditional time convolution structure. The mechanism realizes the timely perception and structural adjustment of the model to the contribution degree change of different modal information, not only improves the modeling ability of the model to complex situations such as load mutation and multi-modal dominant relationship switching, but also enhances the pertinence of feature selection and the flexibility of reasoning path, thereby improving the prediction accuracy, model robustness and generalization ability.

[0110] In the embodiment, the S5 specifically includes:

[0111] S51, receiving a modal routing matrix and performing a multi-threshold segmentation operation thereon to obtain a plurality of levels of connection strength masks and form a connection mask set;

[0112] S52, according to the connection mask set, combining a preset modal channel connection rule, and enumerating a plurality of candidate prediction path structures, each scheme describing a start-stop combination state of a group of modal channels;

[0113] S53, for each candidate prediction path structure, reconstructing its corresponding time convolution propagation path structure, and establishing a structure index table to record its connection topology, active channel and propagation direction information;

[0114] S54, based on each candidate prediction path structure, simulating a one-dimensional convolution propagation process on the original time sequence feature sequence to obtain the corresponding channel output state;

[0115] S55, calculating evaluation indexes of the output state of each candidate prediction path structure, including feature expression completeness, path response strength and structure sparsity, and generating a structure score vector;

[0116] S56, combining each candidate prediction path structure with its corresponding connection scheme, modal routing matrix parameters and structure score vector to form a candidate structure combination.

[0117] The embodiment generates candidate path structures by segmenting the modal routing matrix and constructs a structure score vector, and the system evaluates the expression ability and sparsity of different connection schemes, not only improving the comprehensiveness and discriminability of structure generation, but also providing a high-quality search space for the optimization algorithm part, and enhancing the structure adaptability of prediction.

[0118] In the embodiment, the S6 specifically includes:

[0119] S61, extracting the modal routing matrix, channel connection mode and fusion parameter configuration from the candidate structure combination;

[0120] S62, perform row vector expansion and normalization processing on the modal routing matrix to generate a one-dimensional sparse connection vector representing the connection strength distribution between modes;

[0121] S63, construct a connection index vector for the channel connection mode to record the connection number and timing relationship between each pair of activated channels, and obtain a channel structure code;

[0122] S64, extract the convolution kernel size, connection weight ratio and fusion operation type number from the fusion parameter configuration to generate a fusion parameter code vector;

[0123] S65, concatenate the one-dimensional sparse connection vector, channel structure code and fusion parameter code in a fixed order to generate a structure vector representation;

[0124] S66, combine all structure vector representations to form a structure optimization space.

[0125] The embodiment unifies the encoding of the modal routing matrix, the channel connection mode and the fusion parameter configuration, generates a structure vector representation and constructs a structure optimization space, not only realizes the high-dimensional quantitative expression of the candidate structure, but also improves the comparability and operability of the structure feature, which helps the optimization algorithm to accurately identify the excellent structure combination in the search process, and enhances the optimization efficiency and the global optimality of the predicted structure.

[0126] In the embodiment, the S7 specifically comprises:

[0127] S71, initialize the improved tornado optimization algorithm in the structure optimization space, take the structure vector as the population individual, set the maximum iteration number, population size, upper and lower limits of inertia weight, structure guiding coefficient, disturbance amplitude factor and mutation probability, and evaluate the initial individual with the fitness function, and select the individual with the optimal fitness as the initial center body;

[0128] S72, in each iteration, guide and update each wind body individual according to the structure difference between the center body and the current individual and the structure disturbance direction, and adjust the structure vector using the following formula:

[0129]

[0130] wherein, represents the structure vector of the current individual, represents the center body of the current round, r ％ is the structure guiding coefficient, and γ is the disturbance factor, is the structure disturbance gradient, which is used to enhance the adjustment ability of important structure dimensions;

[0131] S73, dynamically calculate the inertia weight according to the current iteration number, and scale the update result by the inertia coefficient, so that the search step decreases with the iteration, and the inertia coefficient calculation formula is:

[0132]

[0133] w (t) represents the inertia coefficient at the tth iteration, represents the search step control coefficient when the current individual is updated, w ma* represents the initial maximum inertia coefficient, controls the search range at the beginning of the algorithm, w min represents the minimum inertia coefficient, used to limit the search step in the late convergence, T is the maximum iteration number, represents the total number of rounds to be executed in the whole optimization process, t is the current iteration round, and the range is 0≤t≤T;

[0134] S74, perform structure disturbance operation on the individual with a set mutation probability, randomly select another population individual and introduce the structural difference into the current individual, and the update formula is:

[0135]

[0136] wherein, is the structure of the randomly selected individual, and r2 is the mutation amplitude factor, which is used to enhance the diversity of the population and the ability to jump out of the local optimum;

[0137] S75, recalculate the fitness of all updated individuals, and the fitness function is defined as:

[0138] F(x)=α·Strength(x)-β·Sparsity(x)-δ·Error(x);

[0139] wherein, Strength(x) represents the response strength of the structure path, defined as the time average of the weighted sum of the output activation value and the weight of each convolution channel; Sparsity(x) represents the sparsity of the structure connection, calculated as the proportion of the number of unconnected modes in the total number of connections in the modal routing matrix; Error(x) represents the prediction residual, which is the mean square error between the prediction result and the true load value; α, β, δ are the weighting coefficients of each part.

[0140] S76, update the central body structure according to the fitness result, if the maximum iteration number is reached or the optimal structure does not change in continuous multiple iterations, the optimization process is terminated, and the structure vector corresponding to the individual with the optimal fitness is output as the optimal prediction path structure.

[0141] The improved tornado optimization algorithm optimizes the candidate structure combination in the structure optimization space, and contains three innovative mechanisms, including structure guided update, inertia wheel adjustment and random mutation. The structure guided update adjusts the direction based on the difference vector between the current individual and the center body, which improves the global guidance ability of the search process. The inertia wheel adjustment mechanism dynamically controls the search step, optimizes the balance between exploration and convergence. The random mutation mechanism enhances the population diversity by introducing structure disturbance, effectively preventing falling into local optimum. These innovative strategies improve the search efficiency and robustness of the tornado optimization algorithm in complex structure space, so that the final selected prediction path structure has better modal connection configuration and fusion ability, thereby significantly improving the structure adaptability and precision performance of the main network load prediction model.

[0142] In the embodiment, the S8 specifically includes:

[0143] S81, taking the optimal prediction path structure as the structure configuration of the current prediction model, which contains the modal routing matrix, the channel connection mode and the fusion parameter;

[0144] S82, taking the time sequence feature sequence as the prediction model input, performing forward propagation operation according to the channel connection relationship and fusion mode defined by the optimal prediction path structure, and generating the prediction output sequence corresponding to the target time period;

[0145] S83, performing dimension reduction and inverse normalization processing on the prediction output sequence to restore the main network load prediction result consistent with the original load data dimension, and outputting to the target interface according to the set time granularity.

[0146] The embodiment generates the prediction result by using the optimal prediction path structure for model reasoning, and restores it to the actual main network load value. Not only does it ensure the accuracy and physical consistency of the prediction output, but also improves the deployability and interpretability of the model in engineering scenarios, enhancing the practical value.

[0147] Example 1:

[0148] To verify the feasibility of the application in practice, the application is applied to the power load forecasting scenario of a certain high-tech industrial park. The park gathers a number of high-energy-consuming enterprises including precision manufacturing, photovoltaic materials, intelligent equipment, etc. The annual electricity load fluctuates greatly, and the seasonal and behavior-driven characteristics are significant. The conventional prediction model shows large errors after the holiday and during the extreme weather change. In order to improve the prediction accuracy and response ability of the dispatching system, the project team deploys the "a multi-source data fusion prediction method for main network load based on deep learning" proposed by the application to the park energy center dispatching system, and carries out a month-long actual operation test to verify the prediction performance and engineering adaptation ability of the method in the real scene. Then, combined with the key links of load collection, data processing, model structure construction and prediction result analysis, the specific application process and technical advantages of the method in this scene are described in detail.

[0149] At the beginning of the project deployment, the system first accesses and cleans the historical load data provided by the park dispatching master station. The data time resolution is 1 hour, and the data span covers the past 12 months. At the same time, through the local weather interface, temperature, humidity, wind speed, air pressure and other conventional weather variables are collected, and the enterprises in the industrial park upload production calendar, holiday label, peak-valley electricity price response behavior record and other behavior information to form a multi-modal data source input. In the preprocessing stage, the system uses the weighted linear interpolation method to complete 14.6% of the intermittent missing segments, uses the sliding window median detection and range limit method to identify and eliminate 2.3% of the mutant abnormal values, and uses the maximum and minimum value normalization to normalize all modal variables into the [0, 1] interval, ensuring the comparability of the input weight of each data type to the model.

[0150] In the feature construction stage, the system uses multi-type embedding modules to process different modal data. Numerical modalities such as air temperature and electricity load sequence are embedded into a unified dimensional space after two layers of linear transformation, and categorical modalities such as holiday labels are spliced by table lookup encoding and time step position encoding, and then output the same dimension embedding representation after linear mapping. The modal embedding vector is then input into a parallel convolution channel group, which uses different sizes of one-dimensional time convolution kernels for sliding convolution to extract, and constructs three groups of multi-scale convolution channels corresponding to 3 hours, 6 hours and 12 hours sliding window. After the channel output, it is sequentially subjected to batch normalization and ReLU activation layer, and then performs channel dimension splicing operation to form a standardized time series feature sequence as the input of subsequent dynamic modeling.

[0151] In the main prediction structure part, the method constructs an improved timing convolution network based on a dynamic routing control mechanism, generates a modal routing matrix by real-time calculation of channel activation intensity and attention weight. In each time step, according to the connection strength gating strategy, the weak response channel is dynamically closed to form a sparse structure and a response prominent propagation path, and the model realizes rapid structure self-adjustment and stable propagation in the case of rhythm mutation or extreme input. In the candidate path generation stage, the model obtains the candidate connection structure through multi-threshold segmentation and channel combination enumeration, and obtains the feature retention degree, response balance and structure complexity index of each path through simulation propagation and response calculation, to form a candidate structure score vector.

[0152] Subsequently, the structure score vector is input into an improved tornado optimization algorithm module. The algorithm performs multiple rounds of iterative search in the structure vector space, and through the introduction of structure guiding coefficient, dynamic inertia wheel attenuation mechanism and disturbance variation strategy, the balance between local fine-grained search and overall jumping ability is obviously improved. Finally, a set of connection topology optimal, response intensity maximum and parameter sparsity control good structure vector is output as the deployment structure of the actual load prediction model.

[0153] In the actual operation stage of the park, the two weeks from October 14 to October 27, 2024 are selected as the comparison period, and the LSTM model, the GRU model and the method of the application are deployed for prediction test under the same data flow. The evaluation indexes include average MAPE, RMSE, maximum error in peak load period and prediction peak-valley judgment accuracy.

[0154] Table 1 Comparison table of main grid load prediction performance

[0155]

[0156] It can be seen that the method of the application is obviously superior to the traditional model in precision, especially in the scene of load surge caused by the sudden drop of temperature due to the cold wave entering on October 22, the prediction error is controlled within ±400kW; while the error of the traditional LSTM model rises to 750kW, and there is a continuous underestimation problem. The system prediction peak-valley judgment accuracy is improved to 93.9%, and the standby capacity configuration is reduced by about 2.4% compared with the traditional model, and the annual economic savings can reach more than 1.8 million yuan.

[0157] The embodiment verifies the structure plasticity, dynamic adaptability and prediction accuracy advantage of the application through the real data flow and actual deployment of the industrial park, provides a reliable prediction support basis for the high fluctuation area main grid power dispatching system, and has good engineering popularization value.

[0158] The above merely describes preferred specific embodiments of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art, according to the technical solution and inventive concept of the present application, makes equivalent replacement or change within the technical range disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. A deep learning-based main network load multi-source data fusion prediction method, characterized in that, The method comprises the following steps: S1, collecting multi-source input data, and performing time alignment, missing data filling and normalization processing; S2, performing embedding mapping on each modal data to generate modal embedding representation of uniform dimension, and completing feature alignment in time sequence; S3, inputting the modal embedding representation into a convolutional encoder to extract local features of each modal through multi-scale convolution to form a time sequence feature sequence; S4, inputting the time sequence feature sequence into a time sequence convolutional network with a dynamic routing control mechanism, generating a modal routing matrix according to the response mode of each modal channel in the time sequence feature sequence during the convolution modeling process, and dynamically controlling the connection state and propagation path of each modal channel; S5, scheduling the convolutional channel structure based on the modal routing matrix, constructing a candidate prediction path, and forming a corresponding candidate structure combination; S6, constructing a structure optimization space based on the candidate structure, wherein the structure optimization space includes the modal routing matrix, the channel connection mode and the fusion parameter configuration; S7, searching and updating in the structure optimization space by using an improved tornado optimization algorithm to optimize the channel connection mode and the fusion parameter configuration in the candidate structure combination, and determining an optimal prediction path structure; S8, using the optimal prediction path structure for load prediction task to output the main network load prediction result of the target time period.

2. The deep learning-based main network load multi-source data fusion prediction method according to claim 1, characterized in that, The S1 specifically comprises: S11, for samples with time missing, using a weighted linear interpolation algorithm based on the data of adjacent time points to complete the filling, wherein the interpolation result is calculated by weighted average of the front and rear effective values according to the time distance; S12, for abnormal values with mutations, using a combination of sliding median detection and range limitation to remove, wherein the median in the specified length of the sliding window is calculated as the reference, if the current sample deviates from the median more than the set range threshold, the sample is determined as abnormal and replaced with the window median; S13, using the maximum and minimum value normalization method to map each modal data to a unified numerical interval.

3. The deep learning-based main network load multi-source data fusion prediction method according to claim 1, characterized in that, The S2 specifically comprises: S21, identifying the data types of each modal data, marking the continuous variables as numerical type modal data and the discrete variables as category type modal data, and respectively assigning them to the embedding submodules; S22, constructing an embedding structure containing two fully connected networks for numerical type modal data, the first layer uses ReLU activation function, and the second layer maps the original data to a preset uniform embedding dimension space through linear transformation to generate numerical type embedding representation; S23, using a combination of lookup table encoding and position encoding for category type modal data, first mapping the category label to an index vector using a lookup table, then concatenating with the position encoding vector of the corresponding time step, and outputting a category type embedding representation of a specified dimension through a linear layer; S24, performing frame-by-frame alignment on the numerical type embedding representation and the category type embedding representation according to the time index, resampling the embedding sequence with time difference using cubic spline interpolation, and concatenating and fusing each modal embedding representation at each time step to generate modal embedding representation of uniform dimension.

4. The deep learning-based main network load multi-source data fusion prediction method according to claim 1, characterized in that, The S3 specifically comprises: S31, initialize multiple parallel convolution channels of the convolutional encoder, each convolution channel is configured with a time convolution kernel of different size, and the channel output dimension is set to a uniform length; S32, input the modal embedding representation into each convolution channel respectively, perform one-dimensional convolution operation in each channel, and extract sliding feature response at a corresponding scale; S33, sequentially access a batch normalization layer after outputting each convolution channel, and set a ReLU activation function to perform nonlinear mapping processing on the convolution result; S34, splice the processed feature vectors in each convolution channel according to the channel dimension to form a time sequence feature sequence of a unified format.

5. The deep learning-based main network load multi-source data fusion prediction method according to claim 1, characterized in that, The S4 specifically includes: S41, initialize a time sequence convolutional network with a multi-modal channel structure, input the time sequence feature sequence into the corresponding modal channel, and each channel corresponds to an embedded mode one by one; S42, perform one-dimensional convolution operation in each modal channel, extract local activation response features, and calculate the activation mean of the channel at the current time in time steps; S43, form an initial modal attention vector by combining the activation means of each modal channel, and perform Softmax operation to generate normalized modal attention coefficients; S44, perform element-wise multiplication operation on the normalized modal attention coefficients and the corresponding modal gating parameters to form a modal connection strength matrix, and retain high weight items through threshold pruning operation to generate a modal routing matrix; S45, dynamically configure the connection relationship between each modal channel in the time sequence convolutional network according to the modal routing matrix, close the channel path with connection strength lower than the threshold, retain high correlation connection, and realize the structure update of the channel activation state at the current time step; S46, perform time sequence convolution propagation operation on the updated connection structure to complete the dynamic modeling process at the current time step.

6. The deep learning-based main network load multi-source data fusion prediction method according to claim 1, characterized in that, The S5 specifically includes: S51, receive the modal routing matrix and perform multi-threshold segmentation operation to obtain several levels of connection strength masks to form a connection mask set; S52, according to the connection mask set, combine the preset modal channel connection rule to enumerate a plurality of candidate prediction path structures, each scheme describes the on-off combination state of a group of modal channels; S53, for each candidate prediction path structure, reconstruct its corresponding time sequence convolution propagation path structure, and establish a structure index table to record its connection topology, activated channel and propagation direction information; S54, based on each candidate prediction path structure, simulate one-dimensional convolution propagation process on the original time sequence feature sequence to obtain the corresponding channel output state; S55, calculate the evaluation index of the output state of each candidate prediction path structure, including feature expression completeness, path response strength and structure sparsity, and generate a structure score vector; S56, combine each candidate prediction path structure with its corresponding connection scheme, modal routing matrix parameter and structure score vector to form a candidate structure combination.

7. The deep learning-based main network load multi-source data fusion prediction method according to claim 1, characterized in that, The S6 specifically includes: S61, extract the modal routing matrix, channel connection mode and fusion parameter configuration from the candidate structure combination; S62, perform row vector expansion and normalization processing on the modal routing matrix to generate a one-dimensional sparse connection vector representing the connection strength distribution between modes; S63, construct a connection index vector for the channel connection mode to record the connection number and timing relationship between each pair of activated channels, and obtain a channel structure code; S64, extract the convolution kernel size, connection weight ratio, and fusion operation type number from the fusion parameter configuration to generate a fusion parameter code vector; S65, concatenate the one-dimensional sparse connection vector, channel structure code, and fusion parameter code in a fixed order to generate a structure vector representation; S66, combine all structure vector representations to form a structure optimization space. 8.The deep learning-based main network load multi-source data fusion prediction method of claim 1, wherein, The S7 specifically includes: S71, initialize the improved tornado optimization algorithm in the structure optimization space, take the structure vector as the population individual, set the maximum number of iterations, population size, upper and lower limits of inertia weight, structure guide coefficient, disturbance amplitude factor, and mutation probability, and evaluate the initial individual with the fitness function, and select the individual with the optimal fitness as the initial center body; S72, in each iteration, guide the update of each wind body individual according to the structure difference between the center body and the current individual and the structure disturbance direction, and adjust the structure vector using the following formula: wherein, represents the current individual structure vector, represents the current round center body, r ％ is a structure guiding coefficient, γ is a disturbance factor, is a structure disturbance gradient, used to enhance the adjustment ability to important structure dimensions; S73, dynamically calculate the inertia weight according to the current iteration number, and scale the update result according to the inertia coefficient, so that the search step decreases with the iteration, and the inertia coefficient calculation formula is: w (t) denotes the inertia coefficient at the tth iteration, denotes the search step control coefficient when updating the current individual, w ma* denotes the initial maximum inertia coefficient, controls the search range at the initial stage of the algorithm, w min denotes the minimum inertia coefficient, used to limit the search step at the late stage of convergence, T is the maximum number of iterations, denotes the total number of rounds to be performed in the entire optimization process, t is the current iteration round, and the range is 0≤t≤T; S74, perform structure disturbance operation on the individual with the set mutation probability, randomly select another population individual and introduce the structure difference into the current individual, and the update formula is: wherein, r2 is a mutation amplitude factor, this mechanism is used to enhance the diversity of the population and the ability to jump out of local optimum; S75, recalculate the fitness of all updated individuals, and the fitness function is defined as: F(x) = a·Strength(x) - β·Sparsity(x) - δ·Error(x); Wherein, Strength(x) represents the response strength of the structure path, which is defined as the time average of the weighted sum of the output activation value and the weight of each convolution channel; Sparsity(x) represents the sparsity of the structure connection, which is calculated as the proportion of the number of unconnected connections in the total number of connections in the modal routing matrix; Error(x) represents the prediction residual, which is the mean square error between the prediction result and the true load value; a, β, δ are the weighting coefficients of each part; S76, update the center body structure according to the fitness result, if the maximum number of iterations is reached or the optimal structure does not change in continuous multiple iterations, terminate the optimization process, output the structure vector corresponding to the individual with the optimal fitness as the optimal prediction path structure. 9.The deep learning-based main network load multi-source data fusion prediction method of claim 1, wherein, The S8 specifically includes: S81, take the optimal prediction path structure as the structure configuration of the current prediction model, which includes the modal routing matrix, channel connection mode, and fusion parameter; S82, take the time series features as the input of the prediction model, perform forward propagation operation according to the channel connection relationship and fusion mode defined by the optimal prediction path structure, and generate the prediction output sequence corresponding to the target time period; S83, dimension reduction and inverse normalization are performed on the predicted output sequence to restore the main grid load prediction result consistent with the original load data dimension, and output to the target interface according to the set time granularity.

Images