User demand prediction method and system based on multi-modal causal reasoning

Abstract

The invention discloses a user demand prediction method and system based on multi-modal causal reasoning, and the method comprises the steps: carrying out the potential causal relation extraction of a multi-modal data set, and obtaining the causal effect data; constructing an initial long-term demand prediction model based on the causal effect data; training the initial long-term demand prediction model according to the multi-modal data set to obtain a long-term demand prediction model, and decomposing the contribution degree of each user behavior variable to demand intensity based on an SHAP value method in the training process to obtain a causal relationship weight; constructing an initial short-term demand prediction model based on a causal convolutional network based on the causal relationship weight, and inputting the historical short-term user behavior sequence into the initial short-term demand prediction model for training to obtain a short-term demand prediction model; and performing real-time prediction on the user demand according to the short-term demand prediction model and the long-term demand prediction model. According to the method provided by the invention, accurate prediction of user demands is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of demand forecasting, and in particular to a user demand forecasting method and system based on multimodal causal reasoning. Background Art

[0002] As the core digital infrastructure supporting energy transformation, the new energy cloud system carries the critical mission of connecting all aspects of energy production, transmission, and consumption. Within this system, accurate user demand forecasting is crucial for the efficient implementation of energy transformation. The system handles the business operations of a vast number of users and complex environmental events. The dynamic temporal and spatial changes in user demand directly impact the quality of energy services and the efficiency of resource allocation.

[0003] However, existing technologies primarily rely on statistical correlation analysis (such as regression models and time series algorithms) for demand forecasting. This fails to effectively distinguish between true causal relationships and accidental associations between user behavior and environmental events, leading to significant model deviations in scenarios such as policy changes and extreme environmental events. For example, existing technologies attribute "high frequency of user queries after policy releases" to coincidence rather than policy impact, leading to delayed resource scheduling and impacting the reliability of new energy cloud systems. Summary of the Invention

[0004] The present invention provides a user demand prediction method and system based on multimodal causal reasoning to solve the technical problem of how to improve the user demand prediction method of the existing new energy system, thereby achieving the effect of improving the operational reliability of the new energy cloud system.

[0005] To solve the above technical problems, an embodiment of the present invention provides a user demand prediction method based on multimodal causal reasoning, which is applied to a new energy cloud system. The user demand prediction method includes:

[0006] Acquire historical business operation flow data and historical environmental dynamic data of the new energy cloud system, and perform time window alignment and spatial correlation mapping on the historical business operation flow data and the historical environmental dynamic data in sequence to obtain a spatiotemporally aligned multimodal data set;

[0007] Extracting potential causal relationships from the multimodal dataset based on a causal discovery algorithm to obtain causal effect data reflecting the degree of influence of the historical environmental dynamic data on the historical business operation flow data;

[0008] Building an initial long-term demand forecasting model based on the causal effect data, including user behavior, environmental events, and demand intensity;

[0009] Training the initial long-term demand forecasting model according to the multimodal dataset to obtain a long-term demand forecasting model, and decomposing the contribution of each user behavior variable to the demand intensity based on the SHAP value method during the training process to obtain a causal weight;

[0010] Constructing an initial short-term demand forecasting model based on a causal convolutional network based on the causal relationship weights, inputting the acquired historical short-term user behavior sequence of the new energy cloud system into the initial short-term demand forecasting model for training, and obtaining a short-term demand forecasting model;

[0011] In the actual user demand prediction process of the new energy cloud system, user demand is predicted in real time based on the short-term demand prediction model and the long-term demand prediction model.

[0012] As one of the preferred solutions, the historical business operation flow data and the historical environment dynamic data are sequentially subjected to time window alignment and spatial correlation mapping to obtain a spatiotemporally aligned multimodal dataset, including:

[0013] Extracting first timestamp information of each operation behavior in the historical business operation flow data and second timestamp information of each environmental event in the historical environmental dynamic data;

[0014] Time-matching the first timestamp information with the second timestamp information based on a preset time window range to obtain a first multimodal dataset;

[0015] Extracting first geographic location information of each operation behavior in the historical business operation flow data and second geographic location information of each environmental event in the historical environmental dynamic data;

[0016] Based on geographic information system technology, the user operation behavior and the geographical location of environmental events are spatially mapped to obtain the second multimodal dataset;

[0017] The first multimodal dataset and the second multimodal dataset are integrated to obtain a spatiotemporally aligned multimodal dataset.

[0018] As one preferred solution, the causal discovery algorithm is used to extract potential causal relationships from the multimodal dataset to obtain causal effect data reflecting the degree of influence of the historical environmental dynamic data on the historical business operation flow data, including:

[0019] Performing causal structure recognition processing on the multimodal data set based on the FCI algorithm, and constructing a relationship network between variables according to the results of the recognition processing;

[0020] Determine the formation between each variable pair in the relationship network according to the G-test method, and obtain the confounding variable data that affects both the historical environment dynamic data and the historical business operation flow data according to the determination result;

[0021] A directed acyclic graph including key causal edges is constructed based on the confounding variable data, and intervention effect quantification processing is performed on the key causal edges in the directed acyclic graph according to the do-calculus method to obtain the causal effect data.

[0022] As one of the preferred solutions, the SHAP value method is used to decompose the contribution of each user behavior variable to the demand intensity to obtain the causal relationship weight, including:

[0023] According to the SHAP value method, the contribution of each user behavior variable input into the long-term demand forecasting model is decomposed to obtain the average contribution data of each user behavior variable to the demand intensity;

[0024] Each of the average contribution data is normalized to obtain a causal weight reflecting the degree of influence of each user behavior variable on demand intensity.

[0025] As one preferred solution, after real-time prediction of user demand is performed based on the short-term demand prediction model and the long-term demand prediction model, the method further includes:

[0026] Calculating the real-time service load balancing index of the new energy cloud system based on the probability distribution data of users accessing the service progress page output by the short-term demand forecasting model; and preloading the demand corresponding to the high causal weight behavior vector if the real-time service load balancing index is greater than a preset threshold;

[0027] And based on the regional-level horizontal comparison of demand peak period data and hotspot spatiotemporal distribution data output by the long-term demand forecasting model, cross-cycle resource allocation is carried out to pre-generate differentiated service strategies for foreseeable environmental events.

[0028] Another embodiment of the present invention provides a user demand forecasting system based on multimodal causal reasoning, which is applied to a new energy cloud system. The user demand forecasting system includes:

[0029] an acquisition module, configured to acquire historical business operation flow data and historical environmental dynamic data of the new energy cloud system, and sequentially perform time window alignment and spatial correlation mapping on the historical business operation flow data and the historical environmental dynamic data to obtain a spatiotemporally aligned multimodal dataset;

[0030] An extraction module, configured to extract potential causal relationships from the multimodal dataset based on a causal discovery algorithm, and obtain causal effect data reflecting the degree of influence of the historical environmental dynamic data on the historical business operation flow data;

[0031] A first building module is used to build an initial long-term demand forecasting model including user behavior, environmental events and demand intensity based on the causal effect data;

[0032] a training module, configured to train the initial long-term demand forecasting model based on the multimodal dataset to obtain a long-term demand forecasting model, and decompose the contribution of each user behavior variable to the demand intensity based on the SHAP value method during the training process to obtain a causal relationship weight;

[0033] A second construction module is configured to construct an initial short-term demand forecasting model based on a causal convolutional network based on the causal relationship weights, and input the acquired historical short-term user behavior sequence of the new energy cloud system into the initial short-term demand forecasting model for training to obtain a short-term demand forecasting model;

[0034] The prediction module is used to make real-time predictions on user demand based on the short-term demand prediction model and the long-term demand prediction model during the actual user demand prediction process of the new energy cloud system.

[0035] As one preferred solution, the acquisition module is specifically configured to:

[0036] Extracting first timestamp information of each operation behavior in the historical business operation flow data and second timestamp information of each environmental event in the historical environmental dynamic data;

[0037] Time-matching the first timestamp information with the second timestamp information based on a preset time window range to obtain a first multimodal dataset;

[0038] Extracting first geographic location information of each operation behavior in the historical business operation flow data and second geographic location information of each environmental event in the historical environmental dynamic data;

[0039] Based on geographic information system technology, the user operation behavior and the geographical location of environmental events are spatially mapped to obtain the second multimodal dataset;

[0040] The first multimodal dataset and the second multimodal dataset are integrated to obtain a spatiotemporally aligned multimodal dataset.

[0041] As one preferred solution, the extraction module is specifically used to:

[0042] Performing causal structure recognition processing on the multimodal data set based on the FCI algorithm, and constructing a relationship network between variables according to the results of the recognition processing;

[0043] Determine the formation between each variable pair in the relationship network according to the G-test method, and obtain the confounding variable data that affects both the historical environment dynamic data and the historical business operation flow data according to the determination result;

[0044] A directed acyclic graph including key causal edges is constructed based on the confounding variable data, and intervention effect quantification processing is performed on the key causal edges in the directed acyclic graph according to the do-calculus method to obtain the causal effect data.

[0045] As one preferred solution, the training module is further used to:

[0046] According to the SHAP value method, the contribution of each user behavior variable input into the long-term demand forecasting model is decomposed to obtain the average contribution data of each user behavior variable to the demand intensity;

[0047] Each of the average contribution data is normalized to obtain a causal weight reflecting the degree of influence of each user behavior variable on demand intensity.

[0048] As one preferred solution, the prediction module further includes an adjustment unit;

[0049] The adjustment unit is specifically used to:

[0050] Calculating the real-time service load balancing index of the new energy cloud system based on the probability distribution data of users accessing the service progress page output by the short-term demand forecasting model; and preloading the demand corresponding to the high causal weight behavior vector if the real-time service load balancing index is greater than a preset threshold;

[0051] And based on the regional-level horizontal comparison of demand peak period data and hotspot spatiotemporal distribution data output by the long-term demand forecasting model, cross-cycle resource allocation is carried out to pre-generate differentiated service strategies for foreseeable environmental events.

[0052] Compared with the prior art, the embodiments of the present invention have the following advantages:

[0053] 1) This invention uses a causal reasoning mechanism to accurately analyze the true causal relationship between user behavior and demand, effectively distinguishing direct driving factors from indirect related factors, avoiding the causal confusion problem of traditional methods, and significantly improving the accuracy and reliability of demand forecasting.

[0054] 2) The present invention integrates the spatiotemporal features of multimodal data for modeling, which not only captures the short-term dynamic changes and long-term dependencies of user operation sequences, but also combines the impact of environmental events for multi-scale prediction, achieving forward-looking prediction and dynamic response to user needs, and providing efficient support for resource scheduling and service optimization. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 1 is a flow chart of a method for predicting user demand based on multimodal causal reasoning in one embodiment of the present invention;

[0056] Figure 2 is a schematic diagram of a user demand prediction system based on multimodal causal reasoning in one embodiment of the present invention;

[0057] Reference numerals:

[0058] Among them, 11, acquisition module; 12, extraction module; 13, first construction module; 14, training module; 15, second construction module; 16, prediction module. DETAILED DESCRIPTION

[0059] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. The purpose of providing these embodiments is to make the disclosure of the present invention more thorough and comprehensive. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0060] In the description of the present invention, the terms "first," "second," "third," etc. are used for descriptive purposes only and should not be understood to indicate or imply relative importance or implicitly specify the number of the technical features indicated. Therefore, a feature specified as "first," "second," "third," etc. may explicitly or implicitly include one or more of the features. In the description of the present invention, unless otherwise specified, "plurality" means two or more.

[0061] In the description of the present invention, it should be noted that, unless otherwise expressly specified and limited, the terms "installed", "connected" and "connected" should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or an indirect connection through an intermediate medium, or it can be a communication between the two components. The terms "vertical", "horizontal", "left", "right", "up", "down" and similar expressions used herein are for illustrative purposes only, and do not indicate or imply that the device or component referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present invention. The term "and / or" used herein includes any and all combinations of one or more related listed items. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0062] In describing the present invention, it should be noted that, unless otherwise defined, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art. The terms used in the specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. Those skilled in the art will understand the specific meanings of the above terms in the present invention in specific circumstances.

[0063] An embodiment of the present invention provides a method for predicting user demand based on multimodal causal reasoning. For details, see Figure 1 , Figure 1 The figure shows a flow chart of a user demand prediction method based on multimodal causal reasoning in one embodiment of the present invention, which is applied to a new energy cloud system, including steps S1 to S6:

[0064] S1: Acquire historical business operation flow data and historical environmental dynamic data of the new energy cloud system, and perform time window alignment and spatial correlation mapping on the historical business operation flow data and the historical environmental dynamic data in sequence to obtain a spatiotemporally aligned multimodal dataset;

[0065] Business operation flow data includes the time interval for submitting grid connection application materials, the number of views of subsidy application materials, and the version comparison and horizontal comparison behavior. Environmental dynamic data includes regional new energy policy update events and the trigger conditions for power plant operation data comparison requests.

[0066] Specifically, the time interval for submitting grid connection application materials (standard deviation < 30 minutes is marked as urgent) records the entire process from filling in the reserve project information to submitting the grid connection application materials, review feedback, and final grid connection acceptance, with millisecond-level timestamp accuracy. For example, if a user completes the grid connection application submission process within 30 minutes, the application will be marked as "urgent grid connection application processing request."

[0067] Comparing versions of subsidy application materials (e.g., repeated browsing of policy prompts to identify comprehension difficulties) captures details of user interactions on policy documents, technical guides, and power plant data reports, including: Page dwell time: distinguishing between quick browsing (<30 seconds) and in-depth reading (>3 minutes). Cross-document jump paths: For example, if a user jumps from "Photovoltaic Subsidy Policy V2.1" to "Grid Connection Technical Specifications Section 5.2," technical requirements will be prompted.

[0068] The horizontal comparison behavior statistics measure the frequency with which users simultaneously open multiple power plant operation data reports. If a user compares the power generation efficiency, installed capacity, and other indicators of three or more similar power plants within 10 minutes, it is marked as "deep comparison demand."

[0069] Regional new energy policy update events (semantic keyword extraction + release timestamp) use the natural language processing (NLP) engine to parse government announcements and industry notifications in real time, extracting structured information such as policy type (such as subsidy adjustments and technical standard revisions), article keywords, effective time, and impact scope.

[0070] The trigger condition for a power plant operation data comparison request (such as a weekly fluctuation in regional power generation of >5%) is to access the new energy cloud platform to obtain 15-minute granularity output curves to calculate data such as power generation and grid-connected capacity, and calculate the weekly fluctuation rate.

[0071] Preferably, in one embodiment of the present invention, the historical business operation flow data and the historical environment dynamic data are sequentially subjected to time window alignment and spatial correlation mapping processing to obtain a spatiotemporally aligned multimodal dataset, including:

[0072] Extracting first timestamp information of each operation behavior in the historical business operation flow data and second timestamp information of each environmental event in the historical environmental dynamic data;

[0073] Time-matching the first timestamp information with the second timestamp information based on a preset time window range to obtain a first multimodal dataset;

[0074] Extracting first geographic location information of each operation behavior in the historical business operation flow data and second geographic location information of each environmental event in the historical environmental dynamic data;

[0075] Based on geographic information system technology, the user operation behavior and the geographical location of environmental events are spatially mapped to obtain the second multimodal dataset;

[0076] The first multimodal dataset and the second multimodal dataset are integrated to obtain a spatiotemporally aligned multimodal dataset.

[0077] Among them, time window alignment refers to associating the user operation timestamp with the policy release time, and the user behavior timestamp with the policy release time; spatial association mapping refers to aggregating regional-level behavior characteristics based on the user's geographic location.

[0078] It should be noted that business operation flow data and environmental dynamic data naturally have different data formats and spatiotemporal attributes, and need to be aligned to form a unified spatiotemporal coordinate system to provide structured input for subsequent causal analysis and predictive modeling.

[0079] Specifically, in this embodiment, the first timestamp of the user operation is extracted from the business operation flow data. The first timestamp includes the time of submission of the grid connection application, the time of supplementary material submission, etc.; and the second timestamp of the environmental event is extracted from the environmental dynamic data. The second timestamp includes the policy release time, the power generation fluctuation monitoring time, etc. The user behavior timestamp is cross-matched with the environmental event timestamp: if the user visits the relevant page ≥3 times within 1 hour after the policy is released, it is marked as "policy response behavior"; if the power generation in a certain area drops by more than 5% in a single day, the frequency of power station data comparison behavior of users in the area on that day is tracked. A time association label is added to each matched data to form a first multimodal data set, recording features such as the time difference between the behavior and the event, and the response frequency.

[0080] The primary geographic location of user operations is obtained from business operation flow data. This includes registration addresses and power plant access locations. The secondary geographic location of environmental events is extracted from dynamic environmental data. This includes policy implementation areas and power plant cluster locations with fluctuating power generation. Geographic Information System technology is used to coordinate these two types of locations, ensuring that spatial data can be overlaid for analysis.

[0081] User behavior characteristics are aggregated by administrative region or geographic grid to generate regional behavior heat maps. The behavior heat maps are then overlaid with the spatial distribution of new energy facilities to identify the spatial correlation between behavior-intensive areas and energy facilities, forming a second multimodal dataset.

[0082] The time-aligned first multimodal dataset is then associated with the spatially mapped second multimodal dataset to supplement the spatiotemporal cross-features. This ultimately creates a four-dimensional structured dataset containing timestamps, geographic locations, behavior types, and environmental event attributes, providing a unified spatiotemporal input foundation for subsequent causal discovery algorithms and predictive models.

[0083] S2: extracting potential causal relationships from the multimodal dataset based on a causal discovery algorithm to obtain causal effect data reflecting the degree of influence of the historical environmental dynamic data on the historical business operation flow data;

[0084] Preferably, in one embodiment of the present invention, the extracting potential causal relationships from the multimodal dataset based on a causal discovery algorithm to obtain causal effect data reflecting the degree of influence of the historical environmental dynamic data on the historical business operation flow data includes:

[0085] Performing causal structure recognition processing on the multimodal data set based on the FCI algorithm, and constructing a relationship network between variables according to the results of the recognition processing;

[0086] Determine the formation between each variable pair in the relationship network according to the G-test method, and obtain the confounding variable data that affects both the historical environment dynamic data and the historical business operation flow data according to the determination result;

[0087] A directed acyclic graph including key causal edges is constructed based on the confounding variable data, and intervention effect quantification processing is performed on the key causal edges in the directed acyclic graph according to the do-calculus method to obtain the causal effect data.

[0088] The FCI algorithm is a constraint-based causal discovery algorithm that infers the causal structure between variables from observed data through conditional independence tests, outputting a partial directed acyclic graph. G-test is a statistical hypothesis testing method used to determine whether two variables are independent under given conditions and is often used as the core testing tool of the FCI algorithm. Do-calculus is a causal inference tool that calculates causal effects by simulating intervention operations (such as mandatory policy releases).

[0089] In new energy cloud systems, the impact of environmental events on user behavior must be quantified to support resource scheduling. Causal effect data can clarify the specific extent to which policy releases increase subsidy inquiries, avoiding the crudeness of traditional methods that rely solely on empirical thresholds.

[0090] Specifically, in this embodiment, the spatiotemporally aligned multimodal dataset is converted into a behavior-event alignment matrix, which contains user behavior variables, environmental event variables, temporal difference features, and spatial features. The behavior-event alignment matrix is ​​used as input data for the FCI algorithm.

[0091] The FCI algorithm uses conditional independence tests to gradually eliminate irrelevant variables and identify causal-dependency relationships between variables. Specifically, the G-test is used to determine independence between variables, with a confidence threshold of p < 0.01. Based on the output, a partial directed acyclic graph (PDAG) is generated to distinguish between direct causality and indirect correlations. For example, causal relationships such as "policy release → subsidy query ↑" and "power generation fluctuation → data comparison ↑" are clearly identified, and potential confounding variables (such as user-focused energy types such as wind, solar, and storage) are flagged.

[0092] Apply do-calculus to calculate the average causal effect (ACE) for the key causal edges in the directed acyclic graph (such as "policy release → subsidy query"). Specifically expressed as:

[0093] ACE=E[Y|do(X=1)]-E[Y|(X=0)]

[0094] Where X is the intervention variable (e.g., increasing the frequency of subsidy announcements), and Y is the target demand (the number of subsidy inquiries). For example, when the frequency of policy announcements is forced to increase (X = 1), the number of subsidy inquiries (Y) increases by 0.38 times per user (ACE = 0.38, p < 0.001).

[0095] During the parameter estimation phase, the calculated average causal effect (ACE) was used as a constraint in the Neural Structural Equation Model (NeuralSEM). Specifically, for key causal edges (e.g., "policy release → subsidy inquiry"), the quantified result of ACE = 0.38 was used to set the lower limit of the partial derivative of the nonlinear function in the structural equation to 0.38, ensuring that the strength of this causal relationship is prioritized during model parameter training.

[0096] S3: constructing an initial long-term demand forecasting model based on the causal effect data, including user behavior, environmental events, and demand intensity;

[0097] In this embodiment, the observed variables and potential variables are defined, including user behavior (b i ), environmental events (E t ), demand intensity (Y), construct a nonlinear structural equation, and explicitly model the causal relationship between variables. The formula is as follows:

[0098] Y=f(b1,b2,...,b n ,E t )+ξ

[0099] Among them, f is the neural network function and ξ is the error term.

[0100] The causal effect data output from step S2 is directly introduced as a model constraint. Specifically, in the structural equations, for key causal edges, the lower limit of the partial derivative of the neural network weight is set to 0.38 to ensure that the strength of the causal relationship is prioritized during training. Through Bayesian regularization, the confidence level of the causal effect (e.g., p < 0.001) is converted into a weight penalty term to avoid model overfitting and spurious associations.

[0101] S4: training the initial long-term demand forecasting model according to the multimodal dataset to obtain a long-term demand forecasting model, and decomposing the contribution of each user behavior variable to the demand intensity based on the SHAP value method during the training process to obtain a causal relationship weight;

[0102] Preferably, in one embodiment of the present invention, the SHAP value method is used to decompose the contribution of each user behavior variable to the demand intensity to obtain the causal relationship weight, including:

[0103] According to the SHAP value method, the contribution of each user behavior variable input into the long-term demand forecasting model is decomposed to obtain the average contribution data of each user behavior variable to the demand intensity;

[0104] Each of the average contribution data is normalized to obtain a causal weight reflecting the degree of influence of each user behavior variable on demand intensity.

[0105] The gradient descent method is used to optimize the neural network parameters to minimize the prediction error.

[0106] The long-term demand forecasting model (Neural SEM) performs regional-level forecasts by inputting historical user behavior patterns (such as quarterly subsidy inquiry peaks) and planned environmental events (such as technical standard revision announcements). Regional-level forecasts specifically aggregate individual forecast results and output the spatiotemporal distribution of regional demand hotspots.

[0107] Compare the long-term and short-term forecast results with the real-time demand intensity. If there is a significant deviation (such as intensity > 8 but predicted probability < 20%), perform the following operations: data backtracking to check whether the behavior-environment alignment has missed key events (such as failure to capture sudden policy supplements); and model refitting, incrementally train the causal model based on the latest data and update the weight parameters.

[0108] The contribution of each behavioral variable to demand intensity is decomposed using the SHAP (Shapley Additive Explanations) value to obtain the causal weight. The SHAP value method is based on the Shapley value of cooperative game theory and calculates the marginal contribution of each user behavior variable to the demand intensity prediction result.

[0109] Specifically, a spatiotemporally aligned multimodal dataset is input into the long-term demand forecasting model under training, covering different scenarios. For each sample, other variables are fixed, and a certain behavioral variable is changed alone to observe the change in the demand intensity prediction value. The average change value of the variable in all samples is calculated as its SHAP value, which reflects the overall impact of the behavior on demand.

[0110] The SHAP values ​​were converted into causal weights in the range of 0–1 to ensure that behavioral variables of different dimensions could be directly compared. Examples of behavioral types and weights are shown in Table 1.

[0111] Table 1

[0112] Behavior Type Weight range Causal Explanation High-frequency policy inquiries 0.35～0.45 Directly reflects users' sensitivity to policy changes Urgent Project Submission 0.25～0.35 Identify urgent needs in business processes Multi-station data comparison 0.15～0.25 Reflects the user's in-depth analysis intention for performance optimization Technical documentation jumps repeatedly 0.10～0.20 Indicates difficulties in understanding technical details or needs for professional research

[0113] S5: constructing an initial short-term demand forecasting model based on a causal convolutional network based on the causal relationship weights, inputting the acquired historical short-term user behavior sequence of the new energy cloud system into the initial short-term demand forecasting model for training, and obtaining a short-term demand forecasting model;

[0114] After building the long-term demand forecasting model and determining the causal weights, we construct a short-term demand forecasting model. Short-term demand forecasting is crucial for new energy cloud systems, helping them respond promptly to short-term changes in user demand, such as a sudden increase in user inquiries for policy subsidies. By leveraging the previously determined causal weights and building a model based on a causal convolutional network, we can effectively capture the causal information in short-term user behavior sequences, thereby improving the accuracy of short-term demand forecasting.

[0115] The causal weights obtained in step S4 are incorporated into the construction of the causal convolutional network. Specifically, at the network's input layer or convolutional layer, the input features can be weighted according to the causal weights of different user behavior variables. For example, for the behavior variable "high-frequency policy query," its causal weight is high. Therefore, when inputting data, the feature value corresponding to this behavior variable will be given a greater weight, making the model pay more attention to these behavior variables that have a greater impact on demand intensity.

[0116] The input to the initial short-term demand forecasting model is the historical short-term user behavior sequences of the new energy cloud system. These sequences can be pre-processed multimodal data, such as the number of policy queries, project submission frequency, and data comparison operations performed by users over the past few hours or days.

[0117] A causal convolution kernel is used for convolution operations. The convolution kernel design of the Causal Convolutional Network (Causal CNN) is directly related to the causal weights generated in step S4. Specifically, the behavior type weight matrix (e.g., β = 0.45 for high-frequency policy queries) is used as the convolution kernel initialization parameter, and the weight distribution of each channel is dynamically adjusted through the attention mechanism. For example, during the sensitive period after the policy is released, the channel response strength of behavioral features with β > 0.3 is automatically enhanced, while non-causal behavioral features (e.g., β < 0.15) are suppressed.

[0118] The design of the causal convolution kernel ensures that the network only considers the input information at the current moment and before when calculating the output at the current moment. By stacking multiple convolutional layers, features of different scales can be extracted. For example, shallower convolutional layers can extract short-term local features, such as a user's continuous operation patterns over a short period of time; deeper convolutional layers can extract longer-term global features, such as user behavior trends over a period of time.

[0119] Optionally, you can add pooling layers between convolutional layers, such as max pooling or average pooling. The role of the pooling layer is to downsample the feature map, reducing the feature dimension while retaining important feature information, improving the computational efficiency and generalization ability of the model.

[0120] The features extracted by the convolutional and pooling layers are flattened and then mapped to the output space of demand forecasting through a fully connected layer. The fully connected layer can perform nonlinear combinations of features, further improving the model's expressive power.

[0121] The acquired historical short-term user behavior sequences for the new energy cloud system are used as training data. During training, a portion of this data can be used as a validation set to evaluate the model's generalization capabilities. By observing the loss function value and prediction accuracy on the validation set, the model's hyperparameters, such as the convolution kernel size, number of convolution layers, and learning rate, are adjusted to achieve optimal prediction results.

[0122] After training and tuning, the final short-term demand forecasting model was obtained. This model can accurately predict the intensity of short-term user demand based on the historical short-term user behavior sequences of the new energy cloud system. For example, within a short period after a policy is released, the model can predict the changing trend in user inquiries about policy subsidies, providing strong support for the resource scheduling and services of the new energy cloud system.

[0123] S6: In the actual user demand prediction process of the new energy cloud system, real-time prediction of user demand is performed based on the short-term demand prediction model and the long-term demand prediction model.

[0124] In dynamic scheduling strategies, short-term predictions are used to adjust service queue priorities in real time. For example, if the predicted probability of user A accessing the network access review page between 2:00 PM and 3:00 PM is greater than 80%, the system automatically marks the user's request as high priority and preloads the review status summary into the cache.

[0125] Specifically, the service load balancing index is calculated in real time based on the probability distribution of users accessing the business progress page output by the short-term demand forecasting model:

[0126]

[0127] Dynamic scheduling strategy based on service load balancing index:

[0128] Conventional response (L r (t)≤1.0): Process according to the default queue. r When (t)>1.5, a prediction warning is automatically triggered. A robot customer service representative is automatically assigned to handle standard inquiries, freeing up human customer service representatives to handle complex requests.

[0129] In the dynamic scheduling strategy, long-term forecast results are used for regional resource planning. For example, the online customer service resources in the area are increased to 150% of the baseline value three days in advance before the predicted peak demand period (error < 2 hours).

[0130] Specifically, when a user is predicted to inquire about subsidy progress, the system pre-compiles the review status of their application materials, FAQs, and a summary of key policy points. If regional power generation fluctuations trigger a comparison requirement, an analysis report containing the operating data of the top five efficient power plants is automatically generated.

[0131] Preferably, in one embodiment of the present invention, after real-time prediction of user demand is performed according to the short-term demand prediction model and the long-term demand prediction model, the method further includes:

[0132] Calculating the real-time service load balancing index of the new energy cloud system based on the probability distribution data of users accessing the service progress page output by the short-term demand forecasting model; and preloading the demand corresponding to the high causal weight behavior vector if the real-time service load balancing index is greater than a preset threshold;

[0133] And based on the regional-level horizontal comparison of demand peak period data and hotspot spatiotemporal distribution data output by the long-term demand forecasting model, cross-cycle resource allocation is carried out to pre-generate differentiated service strategies for foreseeable environmental events.

[0134] Monitor changes in user behavior after implementation (e.g., number of pre-generated report downloads, increased page dwell time). Calculate service optimization metrics (e.g., percentage reduction in average response time, decrease in repeat query rate). Feedback from monitoring is fed into the incremental learning framework, updating causal and predictive model parameters weekly to adapt to evolving behavioral patterns.

[0135] Specifically, short-term prediction errors (e.g., deviations between predicted probability and actual access time exceeding 15 minutes) are back-propagated to the causal weight calculation module. Using a reinforcement learning algorithm, the causal effect value of a behavior type is dynamically adjusted. For example, if the prediction error for the "repeated jumps to technical documents" behavior remains high, its weight is automatically increased to 0.15-0.25, and the feature extraction layer of the causal convolutional network is retrained.

[0136] An embodiment of the present invention provides a user demand prediction system based on multimodal causal reasoning. For details, see Figure 2 , Figure 2 FIG2 is a schematic diagram of a user demand prediction system based on multimodal causal reasoning in one embodiment of the present invention, which is applied to a new energy cloud system and includes:

[0137] An acquisition module 11 is configured to acquire historical business operation flow data and historical environmental dynamic data of the new energy cloud system, and sequentially perform time window alignment and spatial correlation mapping on the historical business operation flow data and the historical environmental dynamic data to obtain a spatiotemporally aligned multimodal dataset;

[0138] An extraction module 12 is configured to extract potential causal relationships from the multimodal dataset based on a causal discovery algorithm to obtain causal effect data reflecting the degree of influence of the historical environmental dynamic data on the historical business operation flow data;

[0139] A first construction module 13 is configured to construct an initial long-term demand forecasting model including user behavior, environmental events, and demand intensity based on the causal effect data;

[0140] a training module 14 for training the initial long-term demand forecasting model based on the multimodal dataset to obtain a long-term demand forecasting model, and decomposing the contribution of each user behavior variable to the demand intensity based on the SHAP value method during the training process to obtain a causal relationship weight;

[0141] A second construction module 15 is configured to construct an initial short-term demand forecasting model based on a causal convolutional network based on the causal relationship weights, and input the acquired historical short-term user behavior sequence of the new energy cloud system into the initial short-term demand forecasting model for training to obtain a short-term demand forecasting model;

[0142] The prediction module 16 is used to make real-time predictions of user demand based on the short-term demand prediction model and the long-term demand prediction model during the actual user demand prediction process of the new energy cloud system.

[0143] Preferably, in one embodiment of the present invention, the acquisition module 11 is specifically configured to:

[0144] Extracting first timestamp information of each operation behavior in the historical business operation flow data and second timestamp information of each environmental event in the historical environmental dynamic data;

[0145] Time-matching the first timestamp information with the second timestamp information based on a preset time window range to obtain a first multimodal dataset;

[0146] Extracting first geographic location information of each operation behavior in the historical business operation flow data and second geographic location information of each environmental event in the historical environmental dynamic data;

[0147] Based on geographic information system technology, the user operation behavior and the geographical location of environmental events are spatially mapped to obtain the second multimodal dataset;

[0148] The first multimodal dataset and the second multimodal dataset are integrated to obtain a spatiotemporally aligned multimodal dataset.

[0149] Preferably, in one embodiment of the present invention, the extraction module 12 is specifically configured to:

[0150] Performing causal structure recognition processing on the multimodal data set based on the FCI algorithm, and constructing a relationship network between variables according to the results of the recognition processing;

[0151] Determine the formation between each variable pair in the relationship network according to the G-test method, and obtain the confounding variable data that affects both the historical environment dynamic data and the historical business operation flow data according to the determination result;

[0152] A directed acyclic graph including key causal edges is constructed based on the confounding variable data, and intervention effect quantification processing is performed on the key causal edges in the directed acyclic graph according to the do-calculus method to obtain the causal effect data.

[0153] Preferably, in one embodiment of the present invention, the training module 14 is further configured to:

[0154] According to the SHAP value method, the contribution of each user behavior variable input into the long-term demand forecasting model is decomposed to obtain the average contribution data of each user behavior variable to the demand intensity;

[0155] Each of the average contribution data is normalized to obtain a causal weight reflecting the degree of influence of each user behavior variable on demand intensity.

[0156] Preferably, in one embodiment of the present invention, the prediction module 16 further includes an adjustment unit 160;

[0157] The adjustment unit 160 is specifically configured to:

[0158] Calculating the real-time service load balancing index of the new energy cloud system based on the probability distribution data of users accessing the service progress page output by the short-term demand forecasting model; and preloading the demand corresponding to the high causal weight behavior vector if the real-time service load balancing index is greater than a preset threshold;

[0159] And based on the regional-level horizontal comparison of demand peak period data and hotspot spatiotemporal distribution data output by the long-term demand forecasting model, cross-cycle resource allocation is carried out to pre-generate differentiated service strategies for foreseeable environmental events.

[0160] In order to further explain the effectiveness and feasibility of the present invention, the present invention is further described through the following examples.

[0161] Scenario: Demand forecast triggered by revisions to regional photovoltaic subsidy policies.

[0162] In a photovoltaic subsidy policy scenario in a certain area, by using ACE=0.38 as a constraint condition for Neural SEM, the model accurately predicted that the peak of consultation on transitional terms would occur three days in advance, and the regional resource scheduling response speed was 60% faster than traditional methods.

[0163] 1. Data input.

[0164] 1) Environmental data: A certain region released a new policy for local photovoltaic subsidies (effective February 20, 2025), and the number of views of the old policy increased by 220% week-on-week;

[0165] 2) User behavior: User E visited the policy page 7 times within 3 days, 5 of which were to view the comparison between the new and old versions of the article.

[0166] 2. Cause and effect analysis.

[0167] 1) The FCI algorithm identifies key causal chains: policy version update (X) → new and old comparison demand (Y), with confounding variable Z = user-focused energy type wind, solar, and energy storage (p = 0.01 for wind, p = 0.02 for solar, p = 0.03 for energy storage, etc.);

[0168] 2) Calculate demand intensity: S_d = 6.8 (threshold > 5.0), and the attribution report indicates "72% contribution from policy revisions." The intervention effect quantification (ACE), calculated using the do-calculus in step 2.3, serves as a parameter constraint in the neural structural equation model in step 3.2 to ensure causal plausibility in long-term predictions.

[0169] 3. Predict output.

[0170] 1) Short-term: User E has a 91% probability of consulting the transitional terms three days before the policy takes effect. The causal convolutional network directly associates the causal weight matrix generated in step 2 through initialization parameters and the attention mechanism;

[0171] 2) Long term: The region will experience a two-week peak in demand for comparing old and new policies. The neural structural equation model embeds a causal weight matrix in the parameter estimation stage to form structured causal constraints.

[0172] 4. Dynamic response.

[0173] 1) Pre-generate the "Policy Transition Implementation Guide" and distribute it to the intelligent customer service knowledge base;

[0174] 2) Adjust the online customer service resources in this area to 2.5 times the normal level, reducing the response delay to 8 minutes. In the dynamic service optimization strategy in step 4, short-term prediction results are used for real-time scheduling (such as preloading cache), and long-term prediction results are used for resource planning (such as pre-deploying customer service).

[0175] Compared with the prior art, the embodiments of the present invention have the following advantages:

[0176] 1) This invention uses a causal reasoning mechanism to accurately analyze the true causal relationship between user behavior and demand, effectively distinguishing direct driving factors from indirect related factors, avoiding the causal confusion problem of traditional methods, and significantly improving the accuracy and reliability of demand forecasting.

[0177] 2) The present invention integrates the spatiotemporal features of multimodal data for modeling, which not only captures the short-term dynamic changes and long-term dependencies of user operation sequences, but also combines the impact of environmental events for multi-scale prediction, achieving forward-looking prediction and dynamic response to user needs, and providing efficient support for resource scheduling and service optimization.

[0178] The above-described embodiments merely illustrate several implementations of the present invention, and while their descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person skilled in the art would be able to make numerous variations and improvements without departing from the spirit of the present invention, all of which fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be determined by the appended claims.

Claims

1. A user demand prediction method based on multimodal causal reasoning, characterized in that: Applied to the new energy cloud system, the user demand forecasting method includes: Acquire historical business operation flow data and historical environmental dynamic data of the new energy cloud system, and perform time window alignment and spatial correlation mapping on the historical business operation flow data and the historical environmental dynamic data in sequence to obtain a spatiotemporally aligned multimodal data set; Extracting potential causal relationships from the multimodal dataset based on a causal discovery algorithm to obtain causal effect data reflecting the degree of influence of the historical environmental dynamic data on the historical business operation flow data; Building an initial long-term demand forecasting model based on the causal effect data, including user behavior, environmental events, and demand intensity; Training the initial long-term demand forecasting model according to the multimodal dataset to obtain a long-term demand forecasting model, and decomposing the contribution of each user behavior variable to the demand intensity based on the SHAP value method during the training process to obtain a causal weight; Constructing an initial short-term demand forecasting model based on a causal convolutional network based on the causal relationship weights, inputting the acquired historical short-term user behavior sequence of the new energy cloud system into the initial short-term demand forecasting model for training, and obtaining a short-term demand forecasting model; In the actual user demand prediction process of the new energy cloud system, user demand is predicted in real time based on the short-term demand prediction model and the long-term demand prediction model.

2. The user demand prediction method based on multimodal causal reasoning according to claim 1, characterized in that: The step of sequentially performing time window alignment and spatial correlation mapping on the historical business operation flow data and the historical environment dynamic data to obtain a spatiotemporally aligned multimodal dataset includes: Extracting first timestamp information of each operation behavior in the historical business operation flow data and second timestamp information of each environmental event in the historical environmental dynamic data; Time-matching the first timestamp information with the second timestamp information based on a preset time window range to obtain a first multimodal dataset; Extracting first geographic location information of each operation behavior in the historical business operation flow data and second geographic location information of each environmental event in the historical environmental dynamic data; Based on geographic information system technology, the user operation behavior and the geographical location of environmental events are spatially mapped to obtain the second multimodal dataset; The first multimodal dataset and the second multimodal dataset are integrated to obtain a spatiotemporally aligned multimodal dataset.

3. The user demand prediction method based on multimodal causal reasoning according to claim 1, characterized in that: The causal discovery algorithm is used to extract potential causal relationships from the multimodal dataset to obtain causal effect data reflecting the degree of influence of the historical environmental dynamic data on the historical business operation flow data, including: Performing causal structure recognition processing on the multimodal data set based on the FCI algorithm, and constructing a relationship network between variables according to the results of the recognition processing; Determine the formation between each variable pair in the relationship network according to the G-test method, and obtain the confounding variable data that affects both the historical environment dynamic data and the historical business operation flow data according to the determination result; A directed acyclic graph including key causal edges is constructed based on the confounding variable data, and intervention effect quantification processing is performed on the key causal edges in the directed acyclic graph according to the do-calculus method to obtain the causal effect data.

4. The user demand prediction method based on multimodal causal reasoning according to claim 1, characterized in that: The SHAP value method is used to decompose the contribution of each user behavior variable to the demand intensity to obtain the causal relationship weight, including: According to the SHAP value method, the contribution of each user behavior variable input into the long-term demand forecasting model is decomposed to obtain the average contribution data of each user behavior variable to the demand intensity; Each of the average contribution data is normalized to obtain a causal weight reflecting the degree of influence of each user behavior variable on demand intensity.

5. The user demand prediction method based on multimodal causal reasoning according to claim 1, characterized in that: After real-time prediction of user demand is performed according to the short-term demand prediction model and the long-term demand prediction model, the method further includes: Calculating the real-time service load balancing index of the new energy cloud system based on the probability distribution data of users accessing the service progress page output by the short-term demand forecasting model; and preloading the demand corresponding to the high causal weight behavior vector if the real-time service load balancing index is greater than a preset threshold; And based on the regional-level horizontal comparison of demand peak period data and hotspot spatiotemporal distribution data output by the long-term demand forecasting model, cross-cycle resource allocation is carried out to pre-generate differentiated service strategies for foreseeable environmental events.

6. A user demand prediction system based on multimodal causal reasoning, characterized by: Applied to the new energy cloud system, the user demand forecasting system includes: an acquisition module, configured to acquire historical business operation flow data and historical environmental dynamic data of the new energy cloud system, and sequentially perform time window alignment and spatial correlation mapping on the historical business operation flow data and the historical environmental dynamic data to obtain a spatiotemporally aligned multimodal dataset; An extraction module, configured to extract potential causal relationships from the multimodal dataset based on a causal discovery algorithm, and obtain causal effect data reflecting the degree of influence of the historical environmental dynamic data on the historical business operation flow data; A first building module is used to build an initial long-term demand forecasting model including user behavior, environmental events and demand intensity based on the causal effect data; a training module, configured to train the initial long-term demand forecasting model based on the multimodal dataset to obtain a long-term demand forecasting model, and decompose the contribution of each user behavior variable to the demand intensity based on the SHAP value method during the training process to obtain a causal relationship weight; A second construction module is configured to construct an initial short-term demand forecasting model based on a causal convolutional network based on the causal relationship weights, and input the acquired historical short-term user behavior sequence of the new energy cloud system into the initial short-term demand forecasting model for training to obtain a short-term demand forecasting model; The prediction module is used to make real-time predictions on user demand based on the short-term demand prediction model and the long-term demand prediction model during the actual user demand prediction process of the new energy cloud system.

7. The user demand prediction system based on multimodal causal reasoning according to claim 6, characterized in that: The acquisition module is specifically used to: Extracting first timestamp information of each operation behavior in the historical business operation flow data and second timestamp information of each environmental event in the historical environmental dynamic data; Time-matching the first timestamp information with the second timestamp information based on a preset time window range to obtain a first multimodal dataset; Extracting first geographic location information of each operation behavior in the historical business operation flow data and second geographic location information of each environmental event in the historical environmental dynamic data; Based on geographic information system technology, the user operation behavior and the geographical location of environmental events are spatially mapped to obtain the second multimodal dataset; The first multimodal dataset and the second multimodal dataset are integrated to obtain a spatiotemporally aligned multimodal dataset.

8. The user demand forecasting system based on multimodal causal reasoning according to claim 6, characterized in that: The extraction module is specifically used to: performing causal structure recognition processing on the multimodal data set based on the FCI algorithm, and constructing a relationship network between variables according to the results of the recognition processing; Determine the formation between each variable pair in the relationship network according to the G-test method, and obtain the confounding variable data that affects both the historical environment dynamic data and the historical business operation flow data according to the determination result; A directed acyclic graph including key causal edges is constructed based on the confounding variable data, and intervention effect quantification processing is performed on the key causal edges in the directed acyclic graph according to the do-calculus method to obtain the causal effect data.

9. The user demand prediction system based on multimodal causal reasoning according to claim 6, characterized in that: The training module is further used to: According to the SHAP value method, the contribution of each user behavior variable input into the long-term demand forecasting model is decomposed to obtain the average contribution data of each user behavior variable to the demand intensity; Each of the average contribution data is normalized to obtain a causal weight reflecting the degree of influence of each user behavior variable on demand intensity.

10. The user demand prediction system based on multimodal causal reasoning according to claim 6, characterized in that: The prediction module further includes an adjustment unit; The adjustment unit is specifically used to: Calculating the real-time service load balancing index of the new energy cloud system based on the probability distribution data of users accessing the service progress page output by the short-term demand forecasting model; and preloading the demand corresponding to the high causal weight behavior vector if the real-time service load balancing index is greater than a preset threshold; And based on the regional-level horizontal comparison of demand peak period data and hotspot spatiotemporal distribution data output by the long-term demand forecasting model, cross-cycle resource allocation is carried out to pre-generate differentiated service strategies for foreseeable environmental events.

Images