Integrated energy ultra-short-term quantile forecasting method based on dynamic causal perception and conditional density flow

Abstract

This invention discloses a comprehensive energy ultra-short-term quantile prediction method based on dynamic causal perception and conditional density flow, belonging to the field of comprehensive energy system load prediction technology. The method includes: preprocessing collected historical load data of comprehensive energy systems and external meteorological data; establishing a dynamic causal perception mechanism and mapping it to a differentiable soft threshold gating, embedding it into the feature representation learning process in a priori modulation manner to achieve differentiated adjustment of variable contribution under different operating states; constructing a feature-time joint coding structure based on window transfer, effectively capturing local dynamic fluctuations and short-term time-series dependencies of ultra-short-term load through joint attention within the window and cross-window transfer; constructing a conditional normalized flow model, modeling the load probability density based on the conditional features output by the joint coding structure, and deriving quantile prediction results at different confidence levels from the continuous distribution.

Description

Technical Field

[0001] This invention belongs to the field of integrated energy system load forecasting technology, and in particular relates to an integrated energy ultra-short-term quantile forecasting method based on dynamic causal perception and conditional density flow. Background Technology

[0002] Integrated energy systems effectively improve energy utilization efficiency and system operational flexibility by achieving coordinated conversion and optimized operation of multiple energy sources such as electricity, heat, gas, and cooling. Ultra-short-term quantile forecasting of integrated energy load can provide load estimation results at different confidence levels for operation and decision-making levels, serving as a crucial foundation for supporting real-time scheduling, risk-aware operation, and market trading strategy optimization. In ultra-short-term operating scenarios, load uncertainty is particularly prominent. Influenced by rapid weather changes, random user behavior, and the coupling effects of multiple energy sources, integrated energy load exhibits significant random fluctuations on ultra-short-term timescales.

[0003] However, ultra-short-term quantile prediction of integrated energy load still faces several key challenges: First, at ultra-short-term timescales, different energy loads and their interactions with external meteorological factors exhibit significant directional and temporal variations, with the intensity of their influence dynamically evolving with operational status. Existing methods often rely on static correlation analysis or experience-based feature selection strategies, making it difficult to dynamically reflect the relative influence of different variables on load evolution during model training. Second, ultra-short-term load sequences exhibit significant local fluctuations within short time windows, while continuous evolutionary temporal dependencies still exist between adjacent time windows. Traditional time-series modeling methods are insufficient in balancing local dynamic characterization and cross-window dependency modeling. Furthermore, existing quantile prediction methods struggle to accurately characterize complex or even multi-peaked conditional probability distribution structures. Different quantiles are typically estimated independently, lacking consistency constraints at the distribution level, which can easily lead to quantile crossovers and affect the reliability of prediction results. Therefore, there is an urgent need for an ultra-short-term quantile prediction method for integrated energy load that can flexibly introduce structural dependency information during the feature representation stage, effectively model local-global temporal features, and combine highly expressive conditional probability modeling. Summary of the Invention

[0004] The purpose of this invention is to provide a comprehensive energy ultra-short-term quantile prediction method based on dynamic causal perception and conditional density flow, which aims to solve the shortcomings of existing technologies in dynamic correlation characterization of multiple energy variables, collaborative modeling of local and global time series features, and consistency constraints of quantile prediction, thereby improving the reliability and uncertainty characterization capability of load prediction for comprehensive energy systems under ultra-short-term operation scenarios.

[0005] This invention provides a comprehensive energy ultra-short-term quantile prediction method based on dynamic causal sensing and conditional density flow, the method comprising: The collected historical load data of the integrated energy system and external meteorological data are preprocessed; A dynamic causal perception mechanism is established and mapped to a differentiable soft threshold gate. It is embedded into the feature representation learning process in a priori modulation manner, realizing the differentiated adjustment of variable contribution under different operating states. We construct a feature-temporal joint coding structure based on window transfer, and effectively capture the local dynamic fluctuations and short-term dependencies of ultra-short-term loads through joint attention within the window and cross-window transfer. A conditional normalized flow model is constructed, and the load probability density is modeled based on the conditional features output by the joint coding structure. The quantile prediction results at different confidence levels are derived from the continuous distribution.

[0006] Furthermore, the preprocessing of the collected historical load data of the integrated energy system and external meteorological data includes: Data cleaning algorithms are used to identify and remove outliers from the raw data. Missing observations were filled using linear interpolation or time-series mean. Time-scale synchronization of multi-source heterogeneous data is achieved through resampling, and the dimensional unification of each feature variable is completed using the min-max normalization method.

[0007] Furthermore, a dynamic causal perception mechanism is established and mapped to a differentiable soft threshold gating, which is embedded in the feature representation learning process in a priori modulation manner to achieve differentiated adjustment of variable contributions under different operating states, including: Intensity measurement is driven by directional information based on transfer entropy; Dynamic association structure modeling driven by prior directional information; Feature channel gated modulation based on adaptive scale normalization; Learning feature representations embedded in prior modulated signals.

[0008] Furthermore, the intensity metric driven by the directional information based on transfer entropy includes: Calculate the time series of the input variables X With target load time series Y The transfer entropy between Defined as:

[0009] In the formula, Indicates that the input variable is delayed in time. k The state of the step, This indicates the state of the target variable at the previous time step. Given the historical states of the target variable and the historical states of the input variables, the target variable at time [time value missing]. t The conditional probability distribution. This represents a conditional probability distribution that depends solely on the historical state of the target variable itself. Different historical lag steps are set. Obtain directional information driving strength across multiple time scales.

[0010] Furthermore, the dynamic association structure modeling driven by prior directional information includes: The directional information driving strength is mapped to a multivariate correlation weight vector as follows:

[0011] In the formula, M This represents the number of channels in the input variable. Indicates the first m The intensity of directional information corresponding to each input variable.

[0012] Furthermore, the feature channel gating modulation based on adaptive scale normalization includes: Constructing an adaptive scaling normalization factor:

[0013] In the formula, This represents the standard deviation calculation. To prevent small positive numbers with unstable values.

[0014] Construct a continuously differentiable soft threshold gating function based on the normalization factor:

[0015] In the formula, This represents the Sigmoid function. This represents the mean of the weight vector.

[0016] Furthermore, the feature representation learning of the embedded prior modulation signal includes: The prior modulation of the multivariable input feature channel using the soft threshold gating signal is expressed as follows:

[0017] In the formula, This indicates a channel-by-channel product operation. and Let these represent the input feature tensors before and after modulation, respectively, and then convert the modulated input feature tensor... Input subsequent features – time joint modeling module for load forecasting.

[0018] Furthermore, the construction of a feature-temporal joint encoding structure based on window transitions effectively captures the local dynamic fluctuations and short-term dependencies of ultra-short-term loads through in-window joint attention and cross-window transitions, including: Windowed feature representation and local feature-temporal attention modeling; Cross-window transfer and global feature-temporal interaction; Integration of bidirectional long short-term memory networks and fusion of multi-scale features.

[0019] Furthermore, the windowed feature representation and local feature-temporal attention modeling include: Input the multivariate time series as follows:

[0020] In the formula, M Indicates the number of windows. Each window Further divided into N A length of P The total sequence length of the fragment is:

[0021] Through learnable projection matrix Map the original input to D In a 3D feature space, we obtain a window-level feature representation:

[0022] Within each window, a feature-temporal multi-head attention mechanism (WFTMA) is used to model the dependencies between local variables and along the temporal dimension. Within each fragment... h The output of each attention head is:

[0023] In the formula, These are query, key, and value vectors, respectively. This is an element-wise multiplication operation. d For the attention subspace dimension, The dynamic association weight matrix is ​​used to modulate the weights of different input variables in attention calculation.

[0024] Furthermore, the cross-window transfer and global feature-time interaction include: By translating the sequence and performing attention calculations on the translated window, global interaction modeling across windows is achieved. The update process can be represented as follows:

[0025]

[0026]

[0027]

[0028] In the formula, FF is the feedforward network layer, LN is the normalization layer, and it merges the outputs of each window into a single layer. .

[0029] Furthermore, the bidirectional long short-term memory network integration and multi-scale feature fusion include: A bidirectional long short-term memory (BiLSTM) encoder with multiple time scales is constructed to perform parallel modeling of the input sequence and obtain bidirectional encoding results: In the formula, Indicates that the LSTM cell is in t The positive hidden state at any given moment. Indicates that the LSTM cell is in t The hidden state in reverse at any given moment.

[0030] A multi-scale feature fusion strategy is employed to integrate the encoding results from different time scales, forming the final conditional feature representation:

[0031] In the formula, This represents a multi-scale feature fusion operator, including any one of concatenation, weighted summation, or learnable linear mapping. The conditional features are represented... As input to the quantile flow prediction module, it is used to characterize the dynamic evolution of load at different historical dependency scales.

[0032] Furthermore, the introduction of conditional normalized flow models the load probability density under given conditions and consistently derives different quantiles from a continuous distribution, including: Constructing an invertible probability space mapping based on conditional feature constraints; Based on conditional autoregressive structural parameterization characterization of probability density; Calculate the probability density function based on the variable transformation formula; The consistency quantile prediction results are derived based on the probability distribution mechanism.

[0033] Furthermore, the construction of the invertible probability space mapping based on conditional feature constraints includes: Utilizing time t Output conditional feature representation Constructing conditionally invertible mappings :

[0034] In the formula Let the target load be a random variable. To conform to the preset baseline distribution Random variables. Through the conditional mapping function. Map the baseline distribution space to the target load distribution space.

[0035] Furthermore, the probability density based on conditional autoregressive structural parameterization includes: The mapping function is applied using a conditional autoregressive structure. After parameterization, its corresponding conditional probability density is expressed as:

[0036] Each dimension of the conditional distribution is modeled using a Gaussian form:

[0037] In the formula, and These are the conditional mean function and covariance function parameterized by the neural network, respectively.

[0038] Furthermore, the calculation of the probability density function based on the variable transformation formula includes: The inverse mapping Jacobian corresponding to the autoregressive structure is designed as a lower triangular form, and its probability density can be efficiently calculated using the variable transformation formula:

[0039] In the formula, It is a random variable sampled from a uniform distribution. It is the corresponding transformation result. It is the baseline density function. J Let Jacobian matrix represent the inverse mapping.

[0040] Furthermore, the derivation of the consistency quantile prediction result based on the probability distribution mechanism includes: Integrating or inversely calculating the conditional probability density function yields quantile prediction results at different confidence levels; all quantiles are derived using the same conditional probability generation mechanism to eliminate quantile crossover issues and obtain a consistent ultra-short-term load probability prediction distribution.

[0041] This invention provides a device for establishing an ultra-short-term quantile prediction model for integrated energy load, comprising: The data preprocessing module is used to preprocess the collected historical load data of the integrated energy system and external meteorological data; The dynamic causal perception module is used to establish a dynamic causal perception mechanism and map the obtained directional information into a differentiable soft threshold gating signal. It is then embedded into the feature representation learning process in a priori modulation manner to achieve differentiated adjustment of the contribution of input variables under different operating states. The feature-temporal joint coding module is used to construct a feature-temporal joint coding structure based on window transition. Through joint attention within the window and cross-window transition, it effectively captures the local dynamic fluctuations and short-term dependencies of ultra-short-term loads. The quantile prediction module is used to construct a conditionally normalized flow model, model the load probability density based on the conditional features output by the joint coding structure, and derive quantile prediction results at different confidence levels from the continuous distribution.

[0042] An electronic device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the method for establishing an ultra-short-term quantile prediction model for integrated energy load as described above.

[0043] A computer-readable storage medium storing an information transmission implementation program, wherein when executed by a processor, the program implements the steps of the method for establishing an ultra-short-term quantile prediction model for integrated energy load as described above.

[0044] Compared with existing technologies, this invention fully considers the non-stationary fluctuation characteristics, short-term evolution characteristics, and dynamic causal relationships among multiple variables in the ultra-short-term quantile prediction modeling of integrated energy load. By introducing a dynamic causal perception mechanism based on a configurable soft threshold gating, it effectively characterizes the directional dependence of multiple energy loads on external influencing factors over time, and maps the causal structure into a priori modulation signal embedded in the feature learning process, realizing differentiated adjustment of variable contributions and enhancing the model's ability to perceive ultra-short-term sudden fluctuations and rapid state changes. Combined with a feature-time joint coding model based on a window transfer mechanism, it utilizes in-window joint attention and cross-window transfer algorithms to achieve collaborative modeling of local dynamic fluctuations and short-term dependencies, significantly improving the model's ability to express non-stationary interaction features. Simultaneously, it introduces a consistent quantile prediction method based on conditional normalization flow, using normalization flow to accurately model the load probability density under given conditions, and consistently deriving quantiles from continuous conditional distributions, eliminating the quantile crossover problem from a mathematical perspective and ensuring the reliability of uncertainty quantification results in ultra-short-term scenarios. This invention significantly improves the accuracy and robustness of ultra-short-term quantile prediction of comprehensive energy load through an integrated design of dynamic causal perception, feature-time joint modeling, and consistent probability distribution derivation, and has stronger model generalization ability and engineering application value. Attached Figure Description

[0045] Figure 1 This is a flowchart of the method for establishing a comprehensive energy load ultra-short-term quantile prediction model according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the device for establishing the ultra-short-term quantile prediction model of the integrated energy load in Embodiment 1 of the present invention; Figure 3 This is a schematic diagram of the device for establishing the ultra-short-term quantile prediction model of the comprehensive energy load in Embodiment 2 of the present invention. Detailed Implementation

[0046] To overcome the shortcomings of existing technologies, a novel ultra-short-term quantile forecasting method for integrated energy load based on dynamic causal perception representation learning and conditional normalization flow is proposed. This method, through in-depth analysis of the multiple heterogeneous influencing factors in integrated energy load forecasting, constructs an ultra-short-term quantile forecasting model capable of capturing non-stationary fluctuation characteristics, thus achieving an accurate characterization of load uncertainty under complex operating conditions.

[0047] The technical solutions of the embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0048] A comprehensive energy ultra-short-term quantile prediction method based on dynamic causal sensing and conditional density flow is proposed, and its flowchart is shown below. Figure 1 As shown, the specific processing includes the following: Step 1 (S101) involves preprocessing the collected historical load data of the integrated energy system and external meteorological data. Specifically, the following processing methods can be employed: (1) Use data cleaning algorithms to identify and remove outliers from the original data; (2) Use linear interpolation or time series mean to complete the missing observations; (3) Time-scale synchronization of multi-source heterogeneous data is achieved through resampling, and the dimensional uniformity of each feature variable is completed by using the min-max normalization method.

[0049] Step 2 (S102) establishes a dynamic causal perception mechanism and maps it to a differentiable soft threshold gating. This mechanism is embedded into the feature representation learning process using prior modulation, enabling differentiated adjustment of variable contributions under different operating states. Specifically, the following processing can be adopted: Step 2.1, drive the intensity measurement based on the directional information of transfer entropy, the specific process is as follows: Calculate the time series of the input variables X With target load time series Y The transfer entropy between Defined as:

[0050] In the formula, Indicates that the input variable is delayed in time. k The state of the step, This indicates the state of the target variable at the previous time step. Given the historical states of the target variable and the historical states of the input variables, the target variable at time [time value missing]. t The conditional probability distribution. This represents a conditional probability distribution that depends solely on the historical state of the target variable itself. Different historical lag steps are set. Obtain directional information driving strength across multiple time scales.

[0051] Step 2.2, dynamic association structure modeling driven by prior directional information, the specific process is as follows: The directional information driving strength is mapped to a multivariate correlation weight vector as follows:

[0052] In the formula, M This represents the number of channels in the input variable. Indicates the first m The intensity of directional information corresponding to each input variable.

[0053] Step 2.3, feature channel gating modulation based on adaptive scale normalization, the specific process is as follows: Constructing an adaptive scaling normalization factor:

[0054] In the formula, This represents the standard deviation calculation. To prevent small positive numbers with unstable values.

[0055] Construct a continuously differentiable soft threshold gating function based on the normalization factor:

[0056] In the formula, This represents the Sigmoid function. This represents the mean of the weight vector.

[0057] Step 2.4, learning the feature representation of the embedded prior modulation signal, the specific process is as follows: The prior modulation of the multivariable input feature channel using the soft threshold gating signal is expressed as follows:

[0058] In the formula, This indicates a channel-by-channel product operation. and Let these represent the input feature tensors before and after modulation, respectively, and then convert the modulated input feature tensor... Input subsequent features – time joint modeling module for load forecasting.

[0059] Step 3 (S103): Construct a feature-temporal joint encoding structure based on window transition. Through joint attention within the window and cross-window transition, effectively capture the local dynamic fluctuations and short-term temporal dependencies of ultra-short-term loads. Specifically, the following processing can be adopted: Step 3.1, Windowed Feature Representation and Local Feature-Temporal Attention Modeling, the specific process is as follows: Input the multivariate time series as follows:

[0060] In the formula, M Indicates the number of windows. Each window Further divided into N A length of P The total sequence length of the fragment is:

[0061] Through learnable projection matrix Map the original input to D In a 3D feature space, we obtain a window-level feature representation: Within each window, a feature-temporal multi-head attention mechanism (WFTMA) is used to model the dependencies between local variables and along the temporal dimension. h The output of each attention head is:

[0062] In the formula, These are query, key, and value vectors, respectively. This is an element-wise multiplication operation. d For the attention subspace dimension, The dynamic association weight matrix is ​​used to modulate the weights of different input variables in attention calculation.

[0063] Step 3.2, cross-window transfer and global feature-temporal interaction, the specific process is as follows: By translating the sequence and performing attention calculations on the translated window, global interaction modeling across windows is achieved. The update process can be represented as follows:

[0064]

[0065]

[0066]

[0067] In the formula, FF is the feedforward network layer, LN is the normalization layer, and it merges the outputs of each window into a single layer. .

[0068] Step 3.3, integration of bidirectional long short-term memory network and fusion of multi-scale features, the specific process is as follows: A bidirectional long short-term memory (BiLSTM) encoder with multiple time scales is constructed to perform parallel modeling of the input sequence and obtain bidirectional encoding results:

[0069] In the formula, Indicates that the LSTM cell is in t The positive hidden state at any given moment. Indicates that the LSTM cell is in t The hidden state in reverse at any given moment.

[0070] A multi-scale feature fusion strategy is employed to integrate the encoding results from different time scales, forming the final conditional feature representation:

[0071] In the formula, This represents a multi-scale feature fusion operator, including any one of concatenation, weighted summation, or learnable linear mapping. The conditional features are represented... As input to the quantile flow prediction module, it is used to characterize the dynamic evolution of load at different historical dependency scales.

[0072] Step 4 (S104): Construct a conditionally normalized flow model, model the load probability density based on the conditional features output by the joint coding structure, and derive the quantile prediction results at different confidence levels from the continuous distribution. Specifically, the following processing can be adopted: Step 4.1: Construct an invertible mapping in the probability space based on conditional feature constraints. The specific process is as follows: Utilizing time t Output conditional feature representation Constructing conditionally invertible mappings :

[0073] In the formula Let the target load be a random variable. To conform to the preset baseline distribution Random variables. Through the conditional mapping function. Map the baseline distribution space to the target load distribution space.

[0074] Step 4.2, based on the conditional autoregressive structural parameterization, represents the probability density. The specific process is as follows: The mapping function is applied using a conditional autoregressive structure. After parameterization, its corresponding conditional probability density is expressed as:

[0075] Each dimension of the conditional distribution is modeled using a Gaussian form:

[0076] In the formula, and These are the conditional mean function and covariance function parameterized by the neural network, respectively.

[0077] Step 4.3: Calculate the probability density function based on the variable transformation formula. The specific process is as follows: The inverse mapping Jacobian corresponding to the autoregressive structure is designed as a lower triangular form, and its probability density can be efficiently calculated using the variable transformation formula:

[0078] In the formula, It is a random variable sampled from a uniform distribution. It is the corresponding transformation result. It is the baseline density function. J Let Jacobian matrix represent the inverse mapping.

[0079] Step 4.4, derive the consistency quantile prediction results based on the probability distribution mechanism. The specific process is as follows: Integrating or inversely calculating the conditional probability density function yields quantile prediction results at different confidence levels; all quantiles are derived using the same conditional probability generation mechanism to eliminate quantile crossover issues and obtain a consistent ultra-short-term load probability prediction distribution.

[0080] In other words, the specific experimental verification process for the short-term probabilistic prediction of comprehensive energy load is as follows: (1) Setting up the data processing and prediction environment The experimental data in this paper comes from the actual loads (electricity, cooling, and heating) of typical buildings on the ASU Tempe campus in the United States and synchronous meteorological data from the National Climate Data Center (NCDC). The sampling period is from January 2019 to January 2021, with a temporal resolution of 15 minutes. In the data preprocessing stage, linear interpolation was first used to repair outliers and missing values; secondly, normalization was used to eliminate the dimensional differences between multi-source heterogeneous variables to ensure the stability of model convergence. The preprocessed dataset was divided into training, validation, and test sets in a 6:2:2 ratio. This model is built on the PyTorch framework, and the hardware acceleration platform is an NVIDIA RTX 3080Ti GPU.

[0081] (2) Evaluation system for probabilistic prediction performance Unlike deterministic point prediction, this embodiment introduces pinball loss and Winkler score as core evaluation metrics to comprehensively quantify the effectiveness of probabilistic prediction models.

[0082] Pinball loss, as a standardized metric for evaluating the reliability of probabilistic predictions, is calculated using the following formula:

[0083] In the formula, The model represents the first time. Load forecast values ​​output at each quantile This represents the actual observed load value. The smaller this loss fraction, the higher the prediction accuracy and reliability of the model at the corresponding quantile.

[0084] The Winkler score is used to comprehensively consider the coverage and sharpness of the prediction interval. For a given confidence level, its calculation formula is as follows:

[0085] In the formula, The width of the prediction interval (i.e.) ), and These correspond to the upper and lower limits of the prediction interval, respectively. When the actual load value falls within the prediction interval, this indicator is determined solely by the interval width (the narrower the interval, the better); if the actual value exceeds the interval boundary, an additional penalty is triggered. Therefore, the smaller the Winkler score, the better the overall performance of the probability prediction interval.

[0086] (3) Comparative experiment and result analysis To verify the superiority of the proposed probabilistic prediction framework, this study selected several representative baseline models for comparative analysis from two dimensions: the "probability estimation module" and the "temporal feature extraction module." In the probability module, models based on Dynamic Causal Perceptual Representation Learning and Quantile Regression (DCAR-MQ) and Dynamic Causal Perceptual Representation Learning and Gaussian Distribution Hypothesis (DCAR-Gaussian) were selected. In the temporal module, models based on BiLSTM-ATT-CNF (Bidirectional Long Short-Term Memory Network with Attention Mechanism), TCN-CNF (Temporal Convolutional Network with Conditional Normalization Flow), and CNN-BiGRU-ATT-CNF (Convolutional Neural Network with Attention Mechanism, Bidirectional Gated Recurrent Unit with Conditional Normalization Flow) were selected. Using the simplified baseline models and the proposed method (denoted as DCAR-CNF), a probabilistic prediction experiment was conducted on the comprehensive energy load one hour in advance. The quantitative comparison analysis results of the average pinball loss of each model are shown in Table 1, and the evaluation results of the Winkler score at the 90% confidence interval are shown in Table 2.

[0087] Table 1

[0088] Table 2

[0089] In summary, the quantitative results in Tables 1 and 2 show that the proposed short-term probabilistic prediction method for integrated energy load exhibits superior prediction performance in electricity, cooling, and heating scenarios compared to the baseline models. By achieving minimal pinball loss and Winkler score, the proposed method provides a more accurate probability distribution and prediction range, effectively enhancing the risk assessment capability of integrated energy system operation.

[0090] Device Examples According to an embodiment of the present invention, an apparatus for establishing an ultra-short-term quantile prediction model for integrated energy load is provided. Figure 2 This is a schematic diagram of the device for establishing an ultra-short-term quantile prediction model for integrated energy load according to an embodiment of the present invention, as shown below. Figure 2 As shown, the device for establishing a comprehensive energy load ultra-short-term quantile prediction model according to an embodiment of the present invention specifically includes: a data preprocessing module, a dynamic causal sensing module, a feature-time joint encoding module, and a quantile prediction module, thereby obtaining the result of the comprehensive energy load ultra-short-term quantile prediction. Specifically: The data preprocessing module 60 is used to preprocess the collected historical load data of the integrated energy system and external meteorological data. Specifically, the data preprocessing module 60 is used for: Data cleaning algorithms are used to identify and remove outliers from the raw data. Missing observations were filled using linear interpolation or time-series mean. Time-scale synchronization of multi-source heterogeneous data is achieved through resampling, and the dimensional unification of each feature variable is completed using the min-max normalization method.

[0091] The dynamic causal perception module 62 is used to establish a dynamic causal perception mechanism and map the obtained directional information into a differentiable soft threshold gating signal. This signal is then embedded into the feature representation learning process using prior modulation to achieve differentiated adjustment of the contribution of input variables under different operating states. Specifically, the dynamic causal perception module 62 is used for: Intensity measurement is driven by directional information based on transfer entropy; Dynamic association structure modeling driven by prior directional information; Feature channel gated modulation based on adaptive scale normalization; Learning feature representations embedded in prior modulated signals.

[0092] Furthermore, the intensity metric driven by the directional information based on transfer entropy includes: Calculate the time series of the input variables X With target load time series Y The transfer entropy between Defined as:

[0093] In the formula, Indicates that the input variable is delayed in time. k The state of the step, This indicates the state of the target variable at the previous time step. Given the historical states of the target variable and the historical states of the input variables, the target variable at time [time value missing]. t The conditional probability distribution. This represents a conditional probability distribution that depends solely on the historical state of the target variable itself. Different historical lag steps are set. Obtain directional information driving strength across multiple time scales.

[0094] Furthermore, the dynamic association structure modeling driven by prior directional information includes: The directional information driving strength is mapped to a multivariate correlation weight vector as follows:

[0095] In the formula, M This represents the number of channels in the input variable. Indicates the first mThe intensity of directional information corresponding to each input variable.

[0096] Furthermore, the feature channel gating modulation based on adaptive scale normalization includes: Constructing an adaptive scaling normalization factor:

[0097] In the formula, This represents the standard deviation calculation. To prevent small positive numbers with unstable values.

[0098] Construct a continuously differentiable soft threshold gating function based on the normalization factor:

[0099] In the formula, This represents the Sigmoid function. This represents the mean of the weight vector.

[0100] Furthermore, the feature representation learning of the embedded prior modulation signal includes: The prior modulation of the multivariable input feature channel using the soft threshold gating signal is expressed as follows:

[0101] In the formula, This indicates a channel-by-channel product operation. and Let these represent the input feature tensors before and after modulation, respectively, and then convert the modulated input feature tensor... Input subsequent features – time joint modeling module for load forecasting.

[0102] Feature-time joint coding module 62 is used to construct a feature-time joint coding structure based on window transfer. Through in-window joint attention and cross-window transfer, it effectively captures local dynamic fluctuations and short-term temporal dependencies of ultra-short-term loads. Specifically, feature-time joint coding module 62 is used for: Windowed feature representation and local feature-temporal attention modeling; Cross-window transfer and global feature-temporal interaction; Integration of bidirectional long short-term memory networks and fusion of multi-scale features.

[0103] Furthermore, the windowed feature representation and local feature-temporal attention modeling include: Input the multivariate time series as follows:

[0104] In the formula, MIndicates the number of windows. Each window Further divided into N A length of P The total sequence length of the fragment is:

[0105] Through learnable projection matrix Map the original input to D In a 3D feature space, we obtain a window-level feature representation:

[0106] Within each window, a feature-temporal multi-head attention mechanism (WFTMA) is used to model the dependencies between local variables and along the temporal dimension. Within each fragment... h The output of each attention head is:

[0107] In the formula, These are query, key, and value vectors, respectively. This is an element-wise multiplication operation. d For the attention subspace dimension, The dynamic association weight matrix is ​​used to modulate the weights of different input variables in attention calculation.

[0108] Furthermore, the cross-window transfer and global feature-time interaction include: By translating the sequence and performing attention calculations on the translated window, global interaction modeling across windows is achieved. The update process can be represented as follows:

[0109]

[0110]

[0111]

[0112] In the formula, FF is the feedforward network layer, LN is the normalization layer, and it merges the outputs of each window into a single layer. .

[0113] Furthermore, the bidirectional long short-term memory network integration and multi-scale feature fusion include: A bidirectional long short-term memory (BiLSTM) encoder with multiple time scales is constructed to perform parallel modeling of the input sequence and obtain bidirectional encoding results:

[0114] In the formula, Indicates that the LSTM cell is in t The positive hidden state at any given moment. Indicates that the LSTM cell is in t The hidden state in reverse at any given moment.

[0115] A multi-scale feature fusion strategy is employed to integrate the encoding results from different time scales, forming the final conditional feature representation:

[0116] In the formula, This represents a multi-scale feature fusion operator, including any one of concatenation, weighted summation, or learnable linear mapping. The conditional features are represented... As input to the quantile flow prediction module, it is used to characterize the dynamic evolution of load at different historical dependency scales.

[0117] Quantile prediction module 64 is used to construct a conditionally normalized flow model, model the load probability density based on the conditional features output by the joint coding structure, and derive quantile prediction results at different confidence levels from the continuous distribution. Specifically, quantile prediction module 64 is used for: Constructing an invertible probability space mapping based on conditional feature constraints; Based on conditional autoregressive structural parameterization characterization of probability density; Calculate the probability density function based on the variable transformation formula; The consistency quantile prediction results are derived based on the probability distribution mechanism.

[0118] Furthermore, the construction of the invertible probability space mapping based on conditional feature constraints includes: Utilizing time t Output conditional feature representation Constructing conditionally invertible mappings :

[0119] In the formula Let the target load be a random variable. To conform to the preset baseline distribution Random variables. Through the conditional mapping function. Map the baseline distribution space to the target load distribution space.

[0120] Furthermore, the probability density based on conditional autoregressive structural parameterization includes: The mapping function is applied using a conditional autoregressive structure. After parameterization, its corresponding conditional probability density is expressed as:

[0121] Each dimension of the conditional distribution is modeled using a Gaussian form:

[0122] In the formula, and These are the conditional mean function and covariance function parameterized by the neural network, respectively.

[0123] Furthermore, the calculation of the probability density function based on the variable transformation formula includes: The inverse mapping Jacobian corresponding to the autoregressive structure is designed as a lower triangular form, and its probability density can be efficiently calculated using the variable transformation formula:

[0124] In the formula, It is a random variable sampled from a uniform distribution. It is the corresponding transformation result. It is the baseline density function. J Let Jacobian matrix represent the inverse mapping.

[0125] Furthermore, the derivation of the consistency quantile prediction result based on the probability distribution mechanism includes: Integrating or inversely calculating the conditional probability density function yields quantile prediction results at different confidence levels; all quantiles are derived using the same conditional probability generation mechanism to eliminate quantile crossover issues and obtain a consistent ultra-short-term load probability prediction distribution.

[0126] A device for establishing an ultra-short-term quantile prediction model for integrated energy load, see appendix. Figure 2 As shown, it includes: The data preprocessing module 60 is used to preprocess the collected historical load data of the integrated energy system and external meteorological data; The dynamic causal perception module 62 is used to establish a dynamic causal perception mechanism and map the obtained directional information into a differentiable soft threshold gating signal, which is then embedded into the feature representation learning process in a priori modulation manner to achieve differentiated adjustment of the contribution of input variables under different operating states. The feature-temporal joint coding module 64 is used to construct a feature-temporal joint coding structure based on window transition. Through joint attention within the window and cross-window transition, it extracts the local dynamic fluctuations and short-term dependencies of ultra-short-term load. The quantile prediction module 66 is used to construct a conditionally normalized flow model, model the load probability density based on the conditional features output by the joint coding structure, and derive quantile prediction results at different confidence levels from the continuous distribution.

[0127] The present invention also provides an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the above-described method for establishing an ultra-short-term quantile prediction model for integrated energy load.

[0128] The present invention also provides a computer-readable storage medium storing an information transmission implementation program, which, when executed by a processor, implements the steps of the above-described method for establishing an ultra-short-term quantile prediction model for integrated energy load.

[0129] This invention provides an electronic device, such as... Figure 3 As shown, it includes: a memory 70, a processor 72, and a computer program stored in the memory 70 and executable on the processor 72, wherein the computer program, when executed by the processor 72, performs the steps as described in the method embodiment.

[0130] This invention provides a computer-readable storage medium storing an information transmission implementation program, which, when executed by a processor 72, performs the steps described in the method embodiment.

[0131] The computer-readable storage media described in this embodiment include, but are not limited to, ROM, RAM, disk, or optical disk.

Claims

1. A comprehensive energy ultra-short-term quantile prediction method based on dynamic causal perception and conditional density flow, characterized in that, include: S1 preprocesses the collected historical load data of the integrated energy system and external meteorological data; S2 establishes a dynamic causal perception mechanism and maps it to a differentiable soft threshold gating, embedding it into the feature representation learning process in a priori modulation manner to achieve differentiated adjustment of variable contribution under different operating states; S3 constructs a feature-temporal joint encoding structure based on window transfer, which effectively captures the local dynamic fluctuations and short-term dependencies of ultra-short-term loads through joint attention within the window and cross-window transfer. S4 constructs a conditionally normalized flow model, models the load probability density based on the conditional features output by the joint coding structure, and derives quantile prediction results at different confidence levels from the continuous distribution.

2. The integrated energy ultra-short-term quantile prediction method based on dynamic causal perception and conditional density flow as described in claim 1, characterized in that, The specific steps of S2 include: S201 calculates the intensity metric based on the directional information driven by transfer entropy: Calculate the time series of the input variables X With target load time series Y The transfer entropy between Defined as: ， In the formula, Indicates that the input variable is delayed in time. k The state of the step, This represents the state of the target variable at the previous time step. Given the historical states of the target variable and the historical states of the input variables, the target variable at time [time value missing]. t The conditional probability distribution, To determine the conditional probability distribution that depends solely on the historical state of the target variable, different historical lag steps are set. Obtain directional information driving strength across multiple time scales.

3. The integrated energy ultra-short-term quantile prediction method based on dynamic causal perception and conditional density flow as described in claim 2, characterized in that, The specific steps of S2 include: S202 directional information prior-driven dynamic association structure modeling: The directional information driving strength is mapped to a multivariate correlation weight vector as follows: ， In the formula, M This represents the number of channels in the input variable. Indicates the first m The intensity of directional information corresponding to each input variable; S203 is based on adaptive scale-normalized feature channel gated modulation. Constructing an adaptive scaling normalization factor: ， In the formula, This represents the standard deviation calculation. To prevent small positive numbers with unstable values; Construct a continuously differentiable soft threshold gating function based on the aforementioned adaptive scale normalization factor: ， In the formula, This represents the Sigmoid function. This represents the mean of the weight vector; S204 embedding prior modulation signal feature representation learning: The prior modulation of the multivariable input feature channel using the soft threshold gating signal is expressed as follows: ， In the formula, This indicates a channel-by-channel product operation. and Let these represent the input feature tensors before and after modulation, respectively, and then convert the modulated input feature tensor... Input subsequent features – time joint modeling module for load forecasting.

4. The integrated energy ultra-short-term quantile prediction method based on dynamic causal perception and conditional density flow as described in claim 1, characterized in that, The specific steps of S3 include: S301 models windowed feature representations and local feature-temporal attention. Input the multivariate time series as follows: ， In the formula, M Indicates the number of windows, each window Further divided into N A length of P The total sequence length of the fragment is: ， Through learnable projection matrix Map the original input to D In a 3D feature space, we obtain a window-level feature representation: ， Within each window, a feature-temporal multi-head attention mechanism (WFTMA) is used to model the dependencies between local variables and in the temporal dimension. h The output of each attention head is: ， In the formula, These are query, key, and value vectors, respectively. This is an element-wise multiplication operation. d For the attention subspace dimension, The dynamic association weight matrix is ​​used to modulate the weights of different input variables in attention calculation.

5. The integrated energy ultra-short-term quantile prediction method based on dynamic causal perception and conditional density flow according to claim 4, characterized in that, The specific steps of S3 include: S302 performs cross-window transitions and global feature-temporal interactions. This is achieved by shifting the sequence and performing attention calculations on the shifted window, thus modeling global interactions across windows. The update process is represented as follows: ， ， ， ， In the formula, FF is the feedforward network layer, LN is the normalization layer, and it merges the outputs of each window into a single layer. .

6. The integrated energy ultra-short-term quantile prediction method based on dynamic causal perception and conditional density flow as described in claim 5, characterized in that, The specific steps of S3 include: S303 integrates a bidirectional long short-term memory network with multi-scale feature fusion to construct a multi-timescale bidirectional long short-term memory network (BiLSTM) encoder, which performs parallel modeling of the input sequence to obtain bidirectional encoding results. ， In the formula, Indicates that the LSTM cell is in t The positive hidden state at any given moment. Indicates that the LSTM cell is in t The hidden state in reverse at any given moment; A multi-scale feature fusion strategy is employed to integrate the encoding results from different time scales, forming the final conditional feature representation: ， In the formula, The multi-scale feature fusion operator includes any one of concatenation, weighted summation, or learnable linear mapping, which represents the conditional features. As input to the quantile flow prediction module, it is used to characterize the dynamic evolution of load at different historical dependency scales.

7. The integrated energy ultra-short-term quantile prediction method based on dynamic causal perception and conditional density flow as described in claim 1, characterized in that, The specific steps of S4 include: S401 constructs an invertible probability space mapping based on conditional feature constraints, utilizing time... t Output conditional feature representation Constructing conditionally invertible mappings : ， In the formula Let the target load be a random variable. To conform to the preset baseline distribution The random variable, through the conditional mapping function Map the baseline distribution space to the target load distribution space; S402 uses a conditional autoregressive structure to parameterize the probability density and employs the conditional autoregressive structure to represent the mapping function. After parameterization, its corresponding conditional probability density is expressed as: ， Each dimension of the conditional distribution is modeled using a Gaussian form: ， In the formula, and These are the conditional mean function and covariance function parameterized by the neural network, respectively. S403 calculates the probability density function based on the variable transformation formula, and designs the inverse mapping Jacobian corresponding to the autoregressive structure as a lower triangular form, whose probability density can be efficiently calculated using the variable transformation formula: ， In the formula, It is a random variable sampled from a uniform distribution. It is the corresponding transformation result. It is the baseline density function. J The Jacobian matrix representing the inverse mapping; S404 derives consistent quantile prediction results based on the probability distribution mechanism, integrates or reverses the conditional probability density function to obtain quantile prediction results at different confidence levels, and uses the same conditional probability generation mechanism to derive all quantiles to eliminate quantile crossover problems and obtain a consistent ultra-short-term load probability prediction distribution.

8. A device for establishing an ultra-short-term quantile prediction model for integrated energy load, characterized in that, include: The data preprocessing module is used to preprocess the collected historical load data of the integrated energy system and external meteorological data; The dynamic causal perception module is used to establish a dynamic causal perception mechanism and map the obtained directional information into a differentiable soft threshold gating signal. It is then embedded into the feature representation learning process in a priori modulation manner to achieve differentiated adjustment of the contribution of input variables under different operating states. The feature-temporal joint coding module is used to construct a feature-temporal joint coding structure based on window transition. Through joint attention within the window and cross-window transition, it extracts the local dynamic fluctuations and short-term dependencies of ultra-short-term load. The quantile prediction module is used to construct a conditionally normalized flow model, model the load probability density based on the conditional features output by the joint coding structure, and derive quantile prediction results at different confidence levels from the continuous distribution.

9. An electronic device, characterized in that, include: The memory, the processor, and the computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the method for establishing an ultra-short-term quantile prediction model for integrated energy load as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores an information transmission implementation program, which, when executed by a processor, implements the steps of the method for establishing an ultra-short-term quantile prediction model for integrated energy load as described in any one of claims 1 to 7.

Images