Multi-task state joint prediction method and system for user-level short-term load

Abstract

The embodiment of the application provides a kind of for user-level short-term load Multi-task state joint prediction method and system, the method comprises: collecting multi-device operation data, weather data and historical control data, and the multi-device operation data is carried out multidimensional time sequence feature extraction, to generate multidimensional time sequence feature data;The first correlation weight of the multidimensional time sequence feature data, the weather data and the historical control data is calculated to build conversion prediction model;Analysis is carried out between the long-term dependence of the multidimensional time characteristic data, and the multi-state probability condition of the multidimensional time characteristic data is generated;Based on the multi-state probability condition and the output value of the conversion prediction model, build multi-task state joint prediction model, and generate load prediction scheme and equipment operation dynamic adjustable dimension.The embodiment of the application builds multi-modal fusion architecture, realizes the joint output of load value prediction and running state probability distribution.

Description

Technical Field

[0001] This application belongs to the field of integrated energy system forecasting technology, specifically relating to a multi-task state joint forecasting method and system for user-level short-term load. Background Technology

[0002] In the fields of integrated energy systems, virtual power plants, and integrated power generation, grid, load, and storage systems, to achieve optimal system operation economics and maximize the absorption of renewable energy, the formulation of day-ahead dispatch plans requires refined management and control of load resources with adjustment, shifting, or start-up / shutdown capabilities. Therefore, short-term load forecasting technology, as the fundamental support for optimizing day-ahead operation plans, not only needs to accurately predict load values ​​but also dynamically identify load operating states (such as shutdown, maintenance, and operation) to precisely quantify the adjustment margin, shifting range, and control boundaries of loads in different states, thereby ensuring the global optimality of the system operation plan.

[0003] However, existing load forecasting technologies typically focus only on optimizing the accuracy of load forecasts and decouple load forecasting from equipment control strategies. This results in a lack of synergy in operation plan optimization, failing to meet the needs of virtual power plants, integrated energy systems, and integrated power generation, grid, load, and storage systems. The following technical bottlenecks in existing system operation plan optimization urgently need to be overcome: (1) Insufficient correlation of equipment status: The system lacks the ability to intelligently identify the operating status of adjustable loads and cannot automatically identify key parameters. For example, it cannot automatically identify the status of heat pump related equipment (such as whether the heat pump is running, shut down, or under maintenance) or the status of cement plant working equipment (such as whether the cement mill is in start-up, running, shut-down, standby, or under maintenance), which directly affects the assessment of the scheduling potential of controllable resources. (2) Lack of coupling in operation plan optimization: It is difficult to generate optimization instructions for equipment start-up and shutdown timing or power regulation strategy by simply relying on load prediction values. It is necessary to deeply integrate equipment status perception data to realize dynamic adjustment of operation plan.

[0004] While existing technical solutions include methods to correct load forecasts by manually pre-setting the start-up, shutdown, or working state sequence of equipment, this approach has significant drawbacks in terms of load forecast accuracy and applicability. On the one hand, manual experience-based corrections can easily introduce subjective errors, causing the forecast results to deviate from the actual operating conditions. On the other hand, while this method is feasible for small-scale load scenarios, it suffers from low efficiency due to human intervention when the load scale expands. Summary of the Invention

[0005] This application proposes a multi-task state joint prediction method and system for user-level short-term load, which addresses the shortcomings of the prior art.

[0006] According to a first aspect of the embodiments of this application, a multi-task state joint prediction method for user-level short-term load is provided, comprising: Collect multi-device operation data, meteorological data, and historical control data, and extract multi-dimensional time-series features from the multi-device operation data to generate multi-dimensional time-series feature data; Calculate the first correlation weights of the multidimensional time-series feature data, the meteorological data, and the historical control data to build a conversion prediction model; Analyze the long-term dependencies among the multidimensional time feature data and generate the multi-state probability of the multidimensional time feature data; Based on the multi-state probability scenarios and the output values ​​of the transformation prediction model, a multi-task state joint prediction model is built, and a load prediction scheme and dynamically adjustable dimensions of equipment operation are generated.

[0007] In some implementations, the multi-dimensional temporal feature extraction of the multi-device operation data includes: Calculate the statistics of the multi-device operation data within the sliding time window to extract time-domain features; The spectrum is generated by fast Fourier transform, and the frequency band energy ratio and main frequency amplitude features are extracted for frequency domain feature extraction. The periodicity of equipment operating parameters is analyzed by using autocorrelation function, and a difference sequence of adjacent time segments is constructed to characterize the trend of state change, so as to extract time-series correlation features.

[0008] In some implementations, calculating the first association weight of the multidimensional time-series feature data, the meteorological data, and the historical control data includes: Based on the cross-domain correlation of the multidimensional time-series feature data, the meteorological data, and the historical control data at each time step, the correlation value is calculated; The mean of multi-head attention is determined based on the correlation value, and the mean of multi-head attention is determined as the first correlation weight.

[0009] In some implementations, analyzing the long-term dependencies among the multidimensional time feature vectors includes: The multidimensional temporal feature vector is input into a long short-term memory network classification model, and all time steps of the multidimensional temporal feature vector are processed iteratively to capture the long-term dependencies between the features of the multidimensional temporal feature vector.

[0010] In some embodiments, the method further includes: Collect maintenance plan data and electricity price signal data, and cross-domain correlate the maintenance plan data, the electricity price signal data, the multi-equipment operation data, the meteorological data, and the historical control data based on each time step, and calculate the second correlation weight; The conversion prediction model is constructed based on the second association weight.

[0011] In some implementations, the cross-domain feature association of the maintenance plan data, the electricity price signal data, the multi-device operation data, and the historical control data includes: Extract the load fluctuation cycle and abnormal event density from the historical control data; The maintenance plan is mapped to equipment health status labels and aligned with the time sequence of the multi-device operation data. The peak and valley periods of the electricity price signal are analyzed to generate economic weighting coefficients.

[0012] In some implementations, the construction of the multi-task state joint prediction model is expressed by the following formula:

[0013] in, t Indicates time, H Indicates the length of the time interval. This is a joint prediction of the state across multiple tasks. The sequence of the first association weights, These are meteorological factors.

[0014] In some implementations, the objective function of the load forecasting scheme is expressed by the following formula:

[0015] in, This represents the objective function of the load forecasting scheme. The percentage of the predicted value. n This indicates the amount of output data from the joint prediction model. K Indicates the number of device states. S c Indicates that the device is in the first position. c The probability of a class state.

[0016] In some embodiments, the method further includes: Based on the load forecasting scheme and the dynamically adjustable dimensions of equipment operation, real-time operation control data is generated, and the real-time operation control data is input into the processor of the historical control data.

[0017] According to a second aspect of the embodiments of this application, a multi-task state joint prediction system for user-level short-term load is provided, comprising: The feature extraction module is used to collect multi-device operation data, meteorological data and historical control data, and to extract multi-dimensional time-series features from the multi-device operation data to generate multi-dimensional time-series feature data; The conversion prediction building module is used to calculate the first correlation weight of the multidimensional time series feature data, the meteorological data and the historical control data, so as to build a conversion prediction model; The probability analysis module is used to analyze the long-term dependencies between the multidimensional time feature data and generate the multi-state probability of the multidimensional time feature data. The joint prediction module is used to build a multi-task state joint prediction model based on the multi-state probability situation and the output value of the transformation prediction model, and generate a load prediction scheme and a dynamically adjustable dimension of equipment operation.

[0018] The beneficial effects of the multi-task state joint prediction method and system for user-level short-term load in the embodiments of this application include at least the following: This application's embodiments establish a multimodal fusion architecture by integrating equipment state identification and load forecasting through a collaborative modeling mechanism. This addresses the technical deficiency of decoupling load forecasting and equipment control in traditional technologies, achieving joint output of load value prediction and operating state probability distribution. This significantly improves the dynamic optimization capability of operation plans in multi-equipment scenarios (such as integrated energy systems and virtual power plants). Based on a multi-task state joint prediction model, load forecast values ​​and equipment state probabilities are embedded into the same objective function for joint training, overcoming the imbalance between prediction accuracy and scheduling adaptability caused by single-objective optimization. This creates a closed-loop feedback between load forecast results and equipment regulation capability assessment. By outputting load forecast values ​​and equipment state probability distributions in real time, the power adjustment margin and time response boundary of adjustable loads are accurately quantified, adapting to the stringent requirements of integrated energy systems for multi-energy coupling and source-grid-load-storage coordinated scheduling, achieving multi-objective collaborative optimization. When equipment is in an abnormal state (such as a fault shutdown), an adaptive learning mechanism for model parameters is triggered, dynamically correcting the prediction model weight coefficients based on real-time data. This ensures the prediction robustness of the system under non-steady-state conditions, enhancing system applicability and dynamic fault tolerance mechanisms. The modular interface design supports standardized access to photovoltaic output curves, energy storage charging and discharging strategies, and interruptible load control commands, meeting the scalability requirements of large-scale energy resource aggregation scenarios and achieving heterogeneous resource compatibility. Attached Figure Description

[0019] Figure 1 This is a flowchart of an embodiment of the multi-task state joint prediction method for user-level short-term load according to this application; Figure 2This is a flowchart of yet another embodiment of the multi-task state joint prediction method for user-level short-term load according to the present application; Figure 3 This is a schematic diagram of the structure of a multi-task state joint prediction system for short-term user-level load according to an embodiment of this application. Detailed Implementation

[0020] To enable those skilled in the art to better understand the technical solution of this application, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0021] The embodiments of this application will be described in further detail below with reference to the accompanying drawings and examples. The detailed description of the following embodiments and the accompanying drawings are used to illustrate the principles of this application by way of example, but should not be used to limit the scope of this application, that is, this application is not limited to the described embodiments.

[0022] See attached document Figure 1 As shown in the figure, this application discloses specific implementation steps of a multi-task state joint prediction method for user-level short-term load. This method is configured in a multi-task state joint prediction system for user-level short-term load, ensuring that those skilled in the art can implement the technical solution of this application accordingly. The method specifically includes the following steps 110-140.

[0023] Step 110: Collect multi-device operation data, meteorological data, and historical control data, and extract multi-dimensional time-series features from the multi-device operation data to generate multi-dimensional time-series feature data.

[0024] In this context, "multi-device operation data" refers to the operation data of at least one device involved in integrated energy systems, virtual power plants, or integrated power generation, grid, load, and storage systems. Device operation data refers to the physical parameters, status indicators, and performance metrics generated by the device during real-time operation, such as temperature, pressure, current, fault frequency, and energy consumption. Meteorological data characteristics mainly include temperature, humidity, air pressure, wind direction and weather patterns, precipitation, and solar radiation. Historical control data refers to control commands, threshold settings, or intervention records generated by the system or device during historical periods due to policy adjustments and changes in management strategies. The historical control data in this application includes at least time-series data such as transformer historical load rates, temperature rise curves, and voltage regulation commands.

[0025] In some implementations, refer to the appendix. Figure 2As shown, multi-dimensional time-series feature extraction of multi-device operation data includes: inputting the multi-device operation data into the feature extraction engine and calculating the statistics of the multi-device operation data within the sliding time window (such as mean, variance, peak coefficient, and waveform similarity index) for time-domain feature extraction; generating a spectrum through fast Fourier transform and extracting the frequency band energy ratio and main frequency amplitude features for frequency-domain feature extraction; and using the autocorrelation function to analyze the periodicity of the device operation parameters and constructing a difference sequence of adjacent time segments to characterize the state change trend for time-series correlation feature extraction.

[0026] Based on this, the embodiments of this application can combine the focus on the micro-level equipment status of equipment operation data with the macro-level system management of historical control data to achieve real-time perception and dynamic analysis, as well as the mining of policy logic and long-term trends. The synergistic use of the two in energy system optimization not only optimizes system-level decision-making but also improves equipment reliability.

[0027] Step 120: Calculate the first correlation weights of multidimensional time-series feature data, meteorological data, and historical control data to build a conversion prediction model.

[0028] In some implementations, calculating the first association weight of multidimensional time-series feature data, meteorological data, and historical control data includes: cross-domain association of multidimensional time-series feature data, meteorological data, and historical control data based on each time step, and calculating the correlation value; determining the multi-head attention mean based on the correlation value, and determining the multi-head attention mean as the first association weight.

[0029] In some implementations, refer to the appendix. Figure 2 As shown, the embodiments of this application may further include: collecting maintenance plan data and electricity price signal data; cross-domain correlation of maintenance plan data, electricity price signal data, multi-equipment operation data, meteorological data, and historical control data based on each time step; and calculating a second correlation weight; and building a conversion prediction model based on the second correlation weight. The maintenance plan is mainly used to provide discrete event data such as equipment maintenance time and component replacement records to constrain the availability conditions of the prediction model; the electricity price signal is mainly used to introduce time-of-use pricing or demand response pricing, embedding economic objectives into the prediction logic (such as prioritizing maintenance during low-price periods).

[0030] In some implementations, cross-domain feature association of maintenance plan data, electricity price signal data, multi-equipment operation data, and historical control data includes: extracting the load fluctuation cycle (such as time-domain statistics) and abnormal event density (such as overload frequency) from historical control data; mapping maintenance plans to equipment health status labels and aligning them with the time sequence of multi-equipment operation data; and parsing the peak and valley time periods of electricity price signals to generate economic weighting coefficients.

[0031] In one exemplary embodiment, cross-domain association of multidimensional time-series feature data, meteorological data, and historical control data at each time step may include: aligning the multidimensional time-series feature data, meteorological data, and historical control data according to a unified timestamp to ensure consistent data granularity; further, external information such as electricity price signals and equipment health scores may be introduced for association; discrete data may be converted into numerical features through embedding or binning; finally, the interaction term of cross-domain features is calculated, and a neural network is used to assign dynamic weights to features in different domains. Additionally, missing values ​​can be filled using interpolation or the mean of adjacent time windows.

[0032] Among them, the calculation of correlation values ​​can be achieved by parallel computing of multiple attention heads to capture the complex correlations between features in different domains.

[0033] In some implementations, the transformation prediction model is a Transformer prediction model, which is mainly used to map multidimensional features after cross-domain association and calculate the first association weight.

[0034] In addition, the calculation process for the second association weight can be similar to that for the first association weight.

[0035] Step 130: Analyze the long-term dependencies between multidimensional time feature data and generate the multi-state probability of multidimensional time feature data.

[0036] The multiple states can include, for example, "shutdown", "under adjustment" and "available", and load status identification can also use algorithms such as Markov.

[0037] In some implementations, refer to the appendix. Figure 2 As shown, the analysis of long-term dependencies between multidimensional time feature vectors includes: inputting multidimensional time-series feature vectors into a Long Short-Term Memory (LSTM) classification model and iteratively processing all time steps of the multidimensional time-series feature vectors to capture the long-term dependencies between the features of the multidimensional time-series feature vectors.

[0038] In some implementations, the multi-state probability of generating multi-dimensional time feature data also includes multi-state discrimination calculation and multi-state constraint setting.

[0039] The calculation for multi-state discrimination is expressed by the following formula:

[0040] This calculation is based on a classification model using long short-term memory networks, and the device performs calculations at time steps. State characteristics are In the formula, h t Indicates the current time step is tThe hidden state at that time is determined by the LSTM unit based on the current input. X t and the hidden state of the previous moment h t−1 Calculated. Hidden state. h t It is the core of LSTM. h t The transmission of long-term and short-term information of a sequence through memory gating mechanisms (such as forget gate, input gate, and output gate) is the model's "memory" of past information; Indicates that the device is in the first position. i The probability of a class state, For the Softmax function, Meteorological factors, b s These are learnable bias parameters that adjust the model's output distribution in relation to the weight matrix. Used in conjunction with other products.

[0041] The calculation of multi-state constraints is expressed by the following formula: , The device state transition matrix is ​​defined as follows: Limit unreasonable state transitions. ϵ Indicates the noise term. ϵ To follow a normal distribution with a mean of 0 and a standard deviation of 0.1 ( Random noise () represents unpredictable disturbances or errors during state evolution.

[0042] Step 140: Based on the multi-state probability scenarios and the output values ​​of the transformation prediction model, build a multi-task state joint prediction model and generate load prediction schemes and dynamically adjustable dimensions of equipment operation.

[0043] Among them, the multi-task state joint prediction model can predict the load value of multiple devices and the evolution trend of device operation status in the next 24 hours (that is, the dynamic adjustable dimension of device operation).

[0044] In some implementations, the construction of a multi-task state joint prediction model is expressed by the following formula:

[0045] in, t Indicates time, H Indicates the length of the time interval. This is a joint prediction of the state across multiple tasks. The sequence with the first association weight, These are meteorological factors.

[0046] In some implementations, the objective function for generating load forecasting schemes is expressed by the following formula:

[0047] in, This represents the objective function of the load forecasting scheme. The percentage of the predicted value. n This indicates the amount of output data from the joint prediction model. K Indicates the number of device states. S c Indicates that the device is in the first position. c The probability of a class state.

[0048] See attached document Figure 2 As shown, the method also includes: generating real-time operation control data based on load forecasting schemes and the dynamic adjustable dimensions of equipment operation, and inputting the real-time operation control data into a processor for historical control data. The dynamic adjustable dimensions of equipment operation include adjustment space calculation or adjustment potential calculation (e.g., dynamically generating the maximum shiftable duration and remaining charge / discharge cycles of the adjustment space for multiple devices). Generating real-time operation control data is also known as collaborative optimization output, for example, providing a combination of equipment load forecasts and adjustable schemes for scheduling systems such as integrated energy systems, virtual power plants, or microgrids (e.g., "energy storage equipment can discharge 200kW between 15:00 and 16:00").

[0049] This application's embodiments establish a multimodal fusion architecture by integrating equipment state identification and load forecasting through a collaborative modeling mechanism. This addresses the technical deficiency of decoupling load forecasting and equipment control in traditional technologies, achieving joint output of load value prediction and operating state probability distribution. This significantly improves the dynamic optimization capability of operation plans in multi-equipment scenarios (such as integrated energy systems and virtual power plants). Based on a multi-task state joint prediction model, load forecast values ​​and equipment state probabilities are embedded into the same objective function for joint training, overcoming the imbalance between prediction accuracy and scheduling adaptability caused by single-objective optimization. This creates a closed-loop feedback between load forecast results and equipment regulation capability assessment. By outputting load forecast values ​​and equipment state probability distributions in real time, the power adjustment margin and time response boundary of adjustable loads are accurately quantified, adapting to the stringent requirements of integrated energy systems for multi-energy coupling and source-grid-load-storage coordinated scheduling, achieving multi-objective collaborative optimization. When equipment is in an abnormal state (such as a fault shutdown), an adaptive learning mechanism for model parameters is triggered, dynamically correcting the prediction model weight coefficients based on real-time data. This ensures the prediction robustness of the system under non-steady-state conditions, enhancing system applicability and dynamic fault tolerance mechanisms. The modular interface design supports standardized access to photovoltaic output curves, energy storage charging and discharging strategies, and interruptible load control commands, meeting the scalability requirements of large-scale energy resource aggregation scenarios and achieving heterogeneous resource compatibility.

[0050] See attached document Figure 3 As shown in the figure, this application embodiment also provides a multi-task state joint prediction system 300 for user-level short-term load, including: a feature extraction module 310, a transformation prediction construction module 320, a probability analysis module 330 and a joint prediction construction module 340.

[0051] The feature extraction module 310 is used to collect multi-device operation data, meteorological data and historical control data, and to extract multi-dimensional time-series features from the multi-device operation data to generate multi-dimensional time-series feature data.

[0052] The conversion prediction building module 320 is used to calculate the first correlation weights of multidimensional time series feature data, meteorological data and historical control data in order to build a conversion prediction model.

[0053] The probability analysis module 330 is used to analyze the long-term dependencies between multidimensional time feature data and generate multi-state probability cases of multidimensional time feature data.

[0054] The joint forecasting module 340 is used to build a joint forecasting model for multiple task states based on the multi-state probability situation and the output value of the transformation forecasting model, and to generate load forecasting schemes and dynamically adjustable dimensions of equipment operation.

[0055] This application embodiment establishes a multimodal fusion architecture by integrating equipment status identification and load forecasting through a collaborative modeling mechanism. This solves the technical defect of decoupling load forecasting and equipment control in traditional technologies, and realizes the joint output of load value prediction and operating status probability distribution. This significantly improves the dynamic optimization capability of operation plans in multi-equipment scenarios (such as integrated energy systems, virtual power plants, etc.).

[0056] It is understood that the above embodiments are merely exemplary implementations used to illustrate the principles of this application, and this application is not limited thereto. For those skilled in the art, various modifications and improvements can be made without departing from the spirit and substance of this application, and these modifications and improvements are also considered to be within the scope of protection of this application.

Claims

1. A multi-task state joint prediction method for user-level short-term load, characterized in that, include: Collect multi-device operation data, meteorological data, and historical control data, and extract multi-dimensional time-series features from the multi-device operation data to generate multi-dimensional time-series feature data; Calculate the first correlation weights of the multidimensional time-series feature data, the meteorological data, and the historical control data to build a conversion prediction model; Analyze the long-term dependencies among the multidimensional time feature data and generate the multi-state probability of the multidimensional time feature data; Based on the multi-state probability scenarios and the output values ​​of the transformation prediction model, a multi-task state joint prediction model is built, and a load prediction scheme and dynamically adjustable dimensions of equipment operation are generated.

2. The method according to claim 1, characterized in that, The multi-dimensional temporal feature extraction of the multi-device operation data includes: Calculate the statistics of the multi-device operation data within the sliding time window to extract time-domain features; The spectrum is generated by fast Fourier transform, and the frequency band energy ratio and main frequency amplitude features are extracted for frequency domain feature extraction. The periodicity of equipment operating parameters is analyzed by using autocorrelation function, and a difference sequence of adjacent time segments is constructed to characterize the trend of state change, so as to extract time-series correlation features.

3. The method according to claim 2, characterized in that, The calculation of the first correlation weight of the multidimensional time-series feature data, the meteorological data, and the historical control data includes: Based on the cross-domain correlation of the multidimensional time-series feature data, the meteorological data, and the historical control data at each time step, the correlation value is calculated; The mean of multi-head attention is determined based on the correlation value, and the mean of multi-head attention is determined as the first correlation weight.

4. The method according to claim 1, characterized in that, The analysis of the long-term dependencies among the multidimensional time feature vectors includes: The multidimensional temporal feature vector is input into a long short-term memory network classification model, and all time steps of the multidimensional temporal feature vector are processed iteratively to capture the long-term dependencies between the features of the multidimensional temporal feature vector.

5. The method according to claim 1, characterized in that, The method further includes: Collect maintenance plan data and electricity price signal data, and cross-domain correlate the maintenance plan data, the electricity price signal data, the multi-equipment operation data, the meteorological data, and the historical control data based on each time step, and calculate the second correlation weight; The conversion prediction model is constructed based on the second association weight.

6. The method according to claim 5, characterized in that, The cross-domain feature associations of the maintenance plan data, the electricity price signal data, the multi-equipment operation data, and the historical control data include: Extract the load fluctuation cycle and abnormal event density from the historical control data; The maintenance plan is mapped to equipment health status labels and aligned with the time sequence of the multi-device operation data. The peak and valley periods of the electricity price signal are analyzed to generate economic weighting coefficients.

7. The method according to claim 1, characterized in that, The construction of the multi-task state joint prediction model is expressed by the following formula: in, t Indicates time, H Indicates the length of the time interval. This is a joint prediction of the state across multiple tasks. The sequence of the first association weights, These are meteorological factors.

8. The method according to claim 7, characterized in that, The objective function of the load forecasting scheme is expressed by the following formula: in, This represents the objective function of the load forecasting scheme. The percentage of the predicted value. n This indicates the amount of output data from the joint prediction model. K Indicates the number of device states. S c Indicates that the device is in the first position. c The probability of a class state.

9. The method according to claim 1, characterized in that, The method further includes: Based on the load forecasting scheme and the dynamically adjustable dimensions of equipment operation, real-time operation control data is generated, and the real-time operation control data is input into the processor of the historical control data.

10. A multi-task state joint prediction system for user-level short-term load, characterized in that, include: The feature extraction module is used to collect multi-device operation data, meteorological data and historical control data, and to extract multi-dimensional time-series features from the multi-device operation data to generate multi-dimensional time-series feature data; The conversion prediction building module is used to calculate the first correlation weight of the multidimensional time series feature data, the meteorological data and the historical control data, so as to build a conversion prediction model; The probability analysis module is used to analyze the long-term dependencies between the multidimensional time feature data and generate the multi-state probability of the multidimensional time feature data. The joint prediction module is used to build a multi-task state joint prediction model based on the multi-state probability situation and the output value of the transformation prediction model, and generate a load prediction scheme and a dynamically adjustable dimension of equipment operation.

Images