Industrial virtual power plant process event fusion prediction and regulation evaluation method and system

Abstract

The present application belongs to the technical field of virtual power plant resource aggregation and power market dispatching, and particularly relates to an industrial virtual power plant process event fusion prediction and regulation evaluation method and system, which comprises the following steps: collecting multi-source operation data of industrial users, and decomposing total load into basic load and pulse load; inputting the basic load into a multi-scale time sequence convolution network, and performing prediction by fusing an attention gate mechanism; inputting the pulse load into a multi-task learning framework, and obtaining pulse event probability, time offset and peak amplitude through process constraint correction; superimposing the two types of prediction to obtain total load prediction value; defining adjustable state variables and key pulse event locking variables in combination with process state, evaluating user adjustable capacity and aggregating into virtual power plant adjustable capacity boundary, and decomposing dispatching instructions accordingly. The present application realizes accurate prediction of industrial load double mode and quantification of adjustable capacity under process constraint, and provides reliable support for virtual power plant participation in power market transaction.

Description

Technical Field

[0001] This invention belongs to the field of virtual power plant resource aggregation and power market dispatch technology, and particularly relates to a method and system for process event fusion prediction and regulation evaluation in industrial virtual power plants. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] With the advancement of power market reform, virtual power plants have become an important means of aggregating distributed resources and improving system flexibility. Industrial loads, due to their large capacity and strong regularity, have become key aggregation targets.

[0004] However, industrial loads are constrained by production processes and safety boundaries, making it difficult to directly infer their adjustability. In particular, in scenarios such as steel rolling and mine hoisting, the load exhibits a distinct dual-modality: continuous equipment such as fans and pumps form a stable base load; while rolling mills, hoists, and other equipment generate short-term, sudden pulse loads triggered by process events.

[0005] Existing methods predict total load as a single sequence, making it difficult to distinguish between two modes and to predict the timing and peak value of pulse events, thus hindering the determination of whether regulation is permissible. Therefore, there is an urgent need for a process-event fusion prediction and regulation capability assessment method that combines load forecasting, process constraints, pulse event status, and dispatch requirements to support virtual power plants in participating in electricity market transactions. Summary of the Invention

[0006] To overcome the shortcomings of the prior art, this invention provides a method and system for process event fusion prediction and regulation evaluation of industrial virtual power plants. The aim is to establish prediction models for basic load and pulse load by performing dual-modal decomposition of industrial load, and to combine process status and event prediction evaluation to assess adjustability, thereby supporting the safe and accurate aggregation and scheduling of industrial load resources in virtual power plants.

[0007] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions: Firstly, a method for the fusion prediction and adjustment evaluation of process events in industrial virtual power plants is disclosed, including: Collect multi-source data from industrial users, and perform modal decomposition on the total load based on process event signals and load change characteristics to separate the base load and pulse load; The baseline load is input into a multi-scale temporal convolutional network and fused with a multi-attention gating mechanism to obtain the baseline load prediction value. The pulse load is input into a multi-task learning framework to predict the probability of pulse events, the occurrence time offset, and the peak amplitude. After process constraint correction, the pulse load prediction value is obtained. The total load prediction value is obtained by superimposing the baseline load prediction value and the pulse load prediction value. Based on load forecasts and production process status, adjustable state variables and key pulse event locked variables are determined, user adjustability is assessed, and the adjustability of each user is aggregated to obtain the adjustable capacity boundary of the virtual power plant. Dispatch instructions are then decomposed based on the adjustable capacity boundary.

[0008] Furthermore, the collection of multi-source data from industrial users specifically includes: collecting load power data, production process parameters, process event data, equipment status data, production plan data, and virtual power plant scheduling interface data.

[0009] Furthermore, the modal decomposition of the total load includes: when the load increment at adjacent time moments exceeds a preset threshold and a corresponding process event signal exists at the same time, the load component is identified as a pulse load, and the pulse load is separated from the total load to obtain the base load.

[0010] Furthermore, the multi-scale temporal convolutional network contains three parallel branches that capture short-period, medium-period, and long-period features, respectively; the multi-attention gating mechanism includes process attention, periodic attention, and event attention.

[0011] Furthermore, the multi-task learning framework includes a shared feature encoder and three parallel task heads; the shared feature encoder adopts a temporal convolutional network with asymmetric convolutional kernels, and generates pulse feature context vectors by fusing event information with multi-head attention technology; the three task heads are an event detection head for predicting event probability, a time regression head for predicting occurrence time shift, and an amplitude prediction head for predicting peak amplitude.

[0012] Furthermore, the critical pulse event locking variable is used to identify whether there are any uninterruptible or unreducible critical pulse events within a time period.

[0013] Furthermore, the assessment of user adjustability includes: defining process adjustability state variables, which are jointly determined by production process parameters, equipment operating status, and event prediction results, to indicate whether the process status allows load adjustment in the current period; defining critical pulse event lock-in variables to indicate whether there are any uninterrupted or unreducible critical pulse events in the current period; and calculating the user's adjustability based on the above variables. When the process status does not allow adjustment or there is a critical pulse event lock-in, the corresponding adjustability is zero.

[0014] Secondly, a public industrial virtual power plant process event fusion prediction and adjustment evaluation system is provided, including: The data acquisition and decomposition module is used to collect multi-source data from industrial users and perform modal decomposition on the total load based on process event signals and load change characteristics, separating the basic load and the pulse load. The dual-modal load forecasting module is used to input the base load into a multi-scale temporal convolutional network and fuse it with a multi-attention gating mechanism to obtain the base load forecast value; input the pulse load into a multi-task learning framework to predict the probability of pulse events, the occurrence time offset, and the peak amplitude, and correct it with process constraints to obtain the pulse load forecast value; and obtain the total load forecast value by superimposing the base load forecast value and the pulse load forecast value. The adjustment capacity assessment and aggregation scheduling module is used to determine the process adjustable state variables and critical pulse event locked variables based on load forecast values ​​and production process status, assess the user's adjustability, aggregate the adjustability of each user, obtain the adjustable capacity boundary of the virtual power plant, and decompose the scheduling instructions according to the adjustable capacity boundary.

[0015] Thirdly, a computer device is disclosed, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to perform the steps of the method described above.

[0016] Fourthly, a computer-readable storage medium is disclosed having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described above.

[0017] The above one or more technical solutions have the following beneficial effects: This invention decouples industrial load into basic load and pulse load through dual-modal decomposition and differentiated modeling. It employs multi-scale temporal convolutional networks and multi-task learning frameworks for targeted prediction, avoiding the prediction distortion caused by treating pulse load as noise smoothing in traditional methods. At the same time, it integrates a three-head gating mechanism of process attention, cycle attention, and event attention, so that the prediction results take into account the equipment operation rules, production cycle characteristics, and the influence of historical events, significantly improving the accuracy of total load prediction.

[0018] This invention constructs a multi-task learning framework in pulse load prediction that integrates an event detection head, a time regression head, and an amplitude prediction head in parallel and collaborative manner. It also introduces a process-event correlation correction mechanism to achieve accurate three-dimensional prediction of the probability of pulse event occurrence, minute-level time offset, and peak amplitude. This overcomes the shortcomings of existing methods that can only predict load values ​​but cannot predict the timing of events, providing event-level predictive information for production scheduling and helping to optimize equipment scheduling and reduce waiting time.

[0019] This invention deeply integrates load forecasting results with production process status, defines process adjustable state variables and critical pulse event locking variables, and automatically identifies and locks uninterrupted critical production processes when assessing user adjustability, avoiding safety and quality risks caused by virtual power plants erroneously reducing core production loads such as rolling and hoisting. At the same time, it aggregates user-level adjustability into the adjustable capacity boundary of the virtual power plant, supporting the accurate decomposition of dispatch instructions under safety constraints, and providing a safe and quantifiable decision-making basis for virtual power plants to participate in peak shaving, valley filling, demand response, and electricity market application.

[0020] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0021] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0022] Figure 1 This is a flowchart of the industrial virtual power plant process event fusion prediction and regulation evaluation method according to Embodiment 1 of the present invention; Figure 2 This is a flowchart of the basic load prediction process in the industrial virtual power plant process event fusion prediction and regulation evaluation method of Embodiment 1 of the present invention. Figure 3 This is a flowchart of the pulse load prediction process in the industrial virtual power plant process event fusion prediction and regulation evaluation method of Embodiment 1 of the present invention. Figure 4 This is an architecture diagram of the industrial virtual power plant process event fusion prediction and regulation evaluation system according to Embodiment 2 of the present invention. Detailed Implementation

[0023] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0024] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations of the present invention.

[0025] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0026] Example 1 like Figure 1 As shown in the figure, this embodiment discloses a method for process event fusion prediction and regulation evaluation in industrial virtual power plants, including: Step 1: Collect multi-source data from industrial users.

[0027] The industrial users connected to virtual power plants typically include high-energy-consuming industries such as steel, mining, chemicals, building materials, and machinery manufacturing. Multiple types of heterogeneous operational data are collected from various industrial users to comprehensively characterize their load characteristics and production status, specifically including: Load power data, including total incoming power, equipment-level power, feeder power, and power of key production equipment, is used to support subsequent load mode decomposition and predictive modeling. Production process parameters, including temperature, pressure, flow rate, speed, weight, reduction, reaction time, and material state, are used to establish the correlation between load changes and the process. Process event data, including steel biting signals, steel throwing signals, stamping signals, hoisting start signals, arrival stop signals, crushing start signals, etc., are used to identify the triggering timing and type of pulse loads; Equipment status data, including equipment start-up and shutdown status, operating mode, fault status, maintenance status, and controllable status, is used to assess the feasibility and safety boundaries of load regulation. Production planning data, including shift plans, order plans, production line takt time, and equipment scheduling, is used to capture the periodic variation patterns of load. Virtual power plant dispatch interface data, including demand response invitations, market prices, ancillary service instructions, and aggregated dispatch plans, is used to integrate forecast and evaluation results into the dispatch decision-making process.

[0028] After completing the collection of multi-source data, to unify the differences in sampling frequencies among different data sources, all data are time-aligned with a unified time granularity Δt, thereby ensuring temporal consistency in subsequent fusion analysis. Based on this, for the m-th industrial user, the multi-source fusion feature matrix constructed at time t is represented as:

[0029] in, This represents the load power data of industrial users; Represents a vector of production process parameters; This represents the process event state vector. When a process event occurs in time period t, the value is 1; otherwise, it is 0. Represents the device operating state vector; This indicates production planning or process cycle information; This indicates information about virtual power plant scheduling or market constraints.

[0030] Furthermore, to capture the historical dependencies of load changes, the above-mentioned fused features are organized according to time windows to form the following time series input, which serves as the basis for the subsequent prediction model:

[0031] Where L represents the length of the historical observation window. Through the above-mentioned multi-source data acquisition and time-aligned fusion, it is possible to effectively support subsequent dual-modal load decomposition, channel-specific prediction of base load and pulse load, and assessment of adjustability considering process constraints.

[0032] Step 2: Based on the characteristics of process event signals and load mutations, perform modal decomposition on the total load to separate the basic load and the pulse load.

[0033] The total load of industrial users consists of base load and pulse load, and their relationship is as follows:

[0034] in, Indicates the base load; Indicates pulsed load; This indicates unmodeled disturbances or measured noise.

[0035] The two types of loads described above have fundamentally different generation mechanisms and variation patterns. Base loads are mainly generated by continuously operating equipment such as fans, pumps, air compressors, lighting, and control systems. They are characterized by small fluctuations, slow changes, and long-term correlation with the production process, exhibiting strong stability and regularity. In contrast, pulse loads are generated by core production equipment such as rolling mills, stamping mills, elevators, and crushers under the trigger of specific process events. They are characterized by discrete occurrence times, high peak power, short duration, and a clear causal relationship with the process event.

[0036] Based on the aforementioned differences in physical mechanisms, further precise identification of pulse events is performed as a prerequisite for subsequent channel-specific prediction. Specifically, candidate events are first screened based on the temporal abrupt changes in load power. When the load increment at adjacent times satisfies the following formula, a candidate pulse event is determined to exist near that time:

[0037] in, This represents the mutation identification coefficient for the m-th industrial user; This represents the standard deviation of historical load fluctuations. This criterion can effectively capture transient load jumps, but relying solely on power surges is insufficient to distinguish between true process-induced load surges and non-process disturbances.

[0038] Based on the above mutation criteria, this embodiment further introduces process event states. As a confirmation condition, the following real pulse event indicator variables are constructed:

[0039] A pulse is only considered a genuine production pulse load when a significant change in load occurs and a corresponding process event signal exists at the same time. By combining these two conditions, power fluctuations caused by non-process factors such as equipment start-up and shutdown, and measurement noise can be effectively eliminated. This allows for the accurate separation of the pulse load component driven by key processes, providing a precise data foundation for subsequent channel-specific prediction of base load and pulse load.

[0040] Based on this, by separating the identified pulse events from the total load, a continuous and stable base load sequence can be obtained. Thus, the original total load of industrial users is precisely decomposed into two independent components with different physical driving mechanisms, providing a clean and physically meaningful data foundation for the multi-scale time-series prediction of the base load and the event-driven prediction of the impulse load in subsequent steps.

[0041] Step 3: Input the basic load into a multi-scale temporal convolutional network and fuse it with a multi-attention gating mechanism to obtain the predicted basic load value; The base load is influenced by a variety of factors, including production process status, multi-timescale periodicity, slow and gradual changes, and time-varying operating conditions. The difficulty in predicting it lies in simultaneously capturing the dependencies across different time scales and distinguishing the degree of influence of different external factors on load changes. To address this, this invention constructs a multi-scale temporal convolutional network (TCN) as the backbone for basic feature extraction, and introduces a three-head attention mechanism and a gating fusion module to achieve high-precision prediction of the base load.

[0042] First, the input feature vector of the basic load forecasting model is constructed, which includes historical process parameters, equipment status, production plan information, and historical basic load values, specifically represented as follows:

[0043] in, Indicates historical process parameters, Indicates device status. This indicates production planning information. This represents the historical value of the base load.

[0044] Subsequently, a multi-scale temporal convolutional network is used to extract deep features from the above input. This network achieves efficient modeling of historical sequences through dilated causal convolution, and its j-th layer dilated causal convolution is:

[0045] in, This represents the hidden layer output of the j-th layer at time t; The activation function is Sigmoid; The convolution weights of layer j; This represents the drift coefficient of layer j; Indicates the expansion rate Causal convolution operation.

[0046] To capture load variations at different time scales, the temporal convolutional network is configured with three parallel branches, where the short-period branch is used to capture device-level load fluctuations, which can be represented as:

[0047] in, It is the hidden layer output of the short-period branch; It is the short-period part of the multi-scale TCN encoder. Its core parameters include: a kernel size of 3 and an inflation rate of 100%. The experience of the wilderness is 45 steps.

[0048] The intermediate-cycle branch is used to capture load fluctuations at the shift-level cycle, and can be expressed as:

[0049] in, It is the hidden layer output of the medium-period branch; It is the mid-cycle portion of the multi-scale TCN encoder. Its core parameters include: a kernel size of 7 and an inflation rate of 100%. The experience range is 280 steps.

[0050] Long-cycle branches are used for daily / weekly load changes and can be represented as:

[0051] in, It is the hidden layer output of a long-period branch; It is the long-period part of the multi-scale TCN encoder. Its core parameters include: a kernel size of 15 and an inflation rate of 100%. The number of steps taken was 1944.

[0052] Considering the varying degrees and intertwined impacts of process status, time cycles, and historical pulse events on the base load, this embodiment further designs a three-headed attention mechanism based on the aforementioned multi-scale features, to differentiate the influencing factors from the three dimensions of process, cycle, and event: Process attention is used to characterize the impact of the current process state on the base load. The process parameters at the current moment are encoded as a query vector:

[0053] in, A query that indicates the process attention head; The weighting parameters for process attention heads; Let be the vector of process parameters for time period t; For process coding functions; The dimension is The real number field, It is the dimension of the "key" vector.

[0054] After the query is encoded, the key-value pairs are concatenated as follows:

[0055] Where i represents the branch cycle index, we have The elements therein are the tags of the three branches we set for TCN earlier; This represents the key under branch i, indicating the process attention focus. Key weight parameters representing process attention heads; This indicates that the two vectors are concatenated. This represents the t-th step result vector of the hidden layer output of the i-th TCN branch; This represents the process parameters for the t-th time period under the i-th cycle; This represents the value of the process attention point under branch i. This represents the weighting parameter of the process attention head; Indicates the dimension of the value vector.

[0056] Furthermore, the process attention mechanism can be expressed as:

[0057] Therefore, the process context vector under the process attention mechanism is:

[0058] Periodic attention is used to characterize the weighting of load fluctuation patterns over short, medium, and long time periods on base load. The query vector is derived from the multi-scale time-phase encoding at time t.

[0059] in, A query (key) representing periodic attention; The query (key) weight parameter for the periodic attention head; The multi-scale phase encoding at time t is as follows:

[0060] in, , , These correspond to the period lengths of the three branches of TCN, respectively.

[0061] Accordingly, the key-value pairs for periodic attention are:

[0062] Here, the meaning of 'i' is consistent with that of process attention; This represents the predicted output of the i-th branch of the TCN; It is a periodic feature extraction function; Key weight parameters representing the periodic attention head; The value weight parameter represents the periodic attention head; This represents the historical data of the actual industrial base load in the i-th cycle; The encoding function is actually a linear transformation of the real data.

[0063] Since the attention mechanism is the same across all three dimensions, the periodic attention mechanism... This can be expressed as:

[0064] Therefore, the context vector for periodic attention is:

[0065] Event attention is used to assess the residual impact of historical impulse events on the current base load. The query vector is obtained by encoding the event state at the current moment:

[0066] in, The query weight parameter for the event attention header; Encode the event status for time period t (record whether any recent events have occurred, etc.).

[0067] Accordingly, the key-value pair for the event attention head is:

[0068] In the above formula, Key weight parameters for the event attention head; The weight parameter for the event attention head; Encode the event function; This indicates the type of recently occurring event in the i-th branch of the TCN; This represents the load change characteristics (such as pulse amplitude, rise time, etc.) caused by recent events in the i-th branch of the TCN.

[0069] The specific form of the event attention mechanism is represented as follows:

[0070] Therefore, the event context vector is:

[0071] The three attention heads described above model the influencing factors of base load from different perspectives, resulting in context vectors. , , After concatenation along the feature dimension, the input is fused into a gated recurrent unit to obtain a comprehensive context representation:

[0072] in, It is a cyclic gating unit; It is a concatenation of the three-head attention context vectors.

[0073] Finally, the fused context vector is input into a multilayer perceptron to output the future... Step-by-step basic load forecast results:

[0074] in, For the prediction step size, H represents the maximum prediction step size; This represents a multilayer perceptron. It is the transpose of its output layer weight vector. These are the drift parameters of the output layer; This represents the baseline load forecast result at step h after time period t.

[0075] Step 4: Input the pulse load into the multi-task learning framework to predict the probability of pulse events, the occurrence time offset, and the peak amplitude. After process constraint correction, the pulse load prediction value is obtained. In response to the complex characteristics of industrial pulse loads, such as time discreteness, morphological diversity, strong process correlation, and periodic complexity, this embodiment constructs a multi-task learning framework to achieve collaborative prediction of the probability of pulse event occurrence, occurrence time offset, and peak amplitude.

[0076] Industrial pulsed loads exhibit the following complex characteristics: Temporal discreteness, where pulse events are discretely distributed along the time axis with potentially irregular intervals, requiring the model to effectively identify discrete events from continuous time series and predict their precise time points; Morphological diversity, where pulses generated by different equipment exhibit diverse waveforms, including single-peak, multi-peak, flat-top, and spike waveforms, necessitating the simultaneous prediction of multi-dimensional features such as amplitude, width, and rate of rise; Strong process correlation, where pulse events are directly triggered by specific process states (such as coking, rolling, and stamping), requiring the establishment of a causal relationship model from process state to event triggering; and Periodic complexity, where individual equipment has an inherent operating cycle, and multiple equipment working collaboratively generate a compound periodic superposition effect, requiring the model to possess multi-scale periodic modeling capabilities. These four intertwined characteristics significantly distinguish pulsed load prediction from conventional load prediction, necessitating the construction of a specialized processing framework. This embodiment models the above characteristics of industrial pulsed loads as follows: First, the impulse event is defined as the following quintuple:

[0077] in, Indicates the time when the event began; Indicates the time when the event ended; Indicates the peak power of the pulse; Encode the event type (coke pushing, rolling, stamping, etc.); It indicates pulse morphology characteristics, such as rise rate, fall rate, and width ratio.

[0078] To achieve pulse load prediction, this embodiment constructs the following learning framework, as detailed in the appendix. Figure 3 .

[0079] The shared feature encoder in the framework still uses TCN. Traditional TCN convolutional kernels are symmetrical, but pulse shapes are usually asymmetrical, with fast rise and slow fall. Therefore, this example designs three asymmetric convolutional kernels: rising kernel, falling kernel, and flat-top kernel, to enhance the feature response capability to the rising edge, falling edge, and flat-top segment of the pulse, respectively, thereby improving the encoder's accuracy in perceiving pulse shapes. Specifically: Ascending nucleus: = [0.1, 0.3, 0.6, 1.0, 0.8, 0.5]$ (detection pulse rising edge) Descending nucleus: = [1.0, 0.8, 0.5, 0.3, 0.1, 0.0]$ (Detects the falling edge of the pulse) Flat-topped core: = [0.2, 0.5, 0.9, 0.9, 0.5, 0.2]$ (Detecting flat-top pulses) Building upon this, to further integrate multi-source information such as process parameters and event states, this embodiment adds a multi-head attention mechanism to the shared feature encoder:

[0080] in, , , It's about the query, key, and value of this multi-head attention; , , These are the corresponding weight parameters; It is the output of the shared feature encoder.

[0081] The following impulse feature context vector can be obtained through multi-head attention mechanism:

[0082] Where T is the maximum number of iterations; For the attention mechanism at step t; This is the value at step t.

[0083] After obtaining the pulse feature context vector, it is fed together with the output of the shared encoder into the following three parallel task-specific headers to simultaneously complete the multi-dimensional prediction of pulse events.

[0084] The event detection head is used to determine whether a specific type of impulse event has occurred in the current time period. Its network structure takes the impulse context vector as input, passes through two fully connected layers containing 256 and 128 neurons respectively, both using ReLU activation for feature abstraction, and finally maps the output to the (0,1) interval through a single-neuron fully connected layer and a Sigmoid activation function, serving as a probability estimate of the event occurrence. The above process can be formalized as follows:

[0085] in, This represents the output of the fully connected layer (256) in the architecture; and These are its weighting parameters and drift coefficients, respectively; For the output of the fully connected layer (128), and Its weighting parameters and drift coefficients; These are the weight parameters of the fully connected layer (1). Its drift coefficient.

[0086] To ensure that event detection results are consistent with the current process state in terms of physical mechanism, this invention further introduces a process-event correlation correction mechanism after the event detection head. Specifically, it first dynamically calculates the correlation weight between the current process parameter vector and the event probabilities at each historical time point through an attention mechanism:

[0087] in, Represents the transpose of the process vector; Here are the weight parameters; the subscript j indicates the time step.

[0088] Then, the process coding features at each time point are weighted and fused to obtain the process-event related context vector:

[0089] Where L represents the process vector The total length.

[0090] Based on this, the current process code is concatenated with the aforementioned context vector and then sent to the gate control loop unit to generate a corrected gate control signal:

[0091] in, This represents a gated loop unit; and These are its weight parameters and drift parameters, respectively.

[0092] Finally, the original event probabilities are weighted and modulated using this gated signal to obtain the event probabilities corrected for process constraints:

[0093] The time regression head is responsible for predicting the precise time offset of impulse events. Its network structure also takes the impulse context vector as input, passes through two fully connected layers (256-dimensional and 128-dimensional, ReLU activated), and then uses a single-neuron fully connected layer with a Tanh activation function to restrict the output to the [-1, 1] interval, before multiplying it by the maximum allowable delay constant. The relative time offset is obtained:

[0094] Given maximum allowable delay (For example, 15 minutes), the predicted time of the event is as follows:

[0095] The amplitude prediction head is used to predict the peak power of impulse events. Its network structure also takes the impulse context vector as input, extracts features through two fully connected layers (256-dimensional and 128-dimensional, ReLU activated), and finally uses a single-neuron fully connected layer with the Softplus activation function to ensure the output is always positive, thus obtaining the impulse amplitude prediction result.

[0096] After the above parallel processing and correction, this embodiment finally obtains the three core elements in the pulse event quintuple. event probability , With peak amplitude This provides complete pulse-level prediction information for subsequent pulse load reconfiguration and virtual power plant dispatching decisions.

[0097] Step 5: Obtain the total load forecast value by superimposing the basic load forecast value and the pulse load forecast value; Based on the baseline load forecast obtained in step 3 and the pulse load forecast obtained in step 4, the two are superimposed to obtain the total load forecast. ):

[0098] in, It is a pulse indicator function, specifically:

[0099] in, The prediction results can be obtained through the time prediction head. We can deduce this by working backwards.

[0100] To improve prediction accuracy, a method is constructed that includes event detection loss. Time regression loss Amplitude prediction loss Process constraint losses Coordinated scheduling constraint loss And consistency constraint loss Overall loss function Train the model:

[0101] in, This is the overall loss function. The specific meanings and forms of the remaining terms are as follows: The event detection loss function adopts the form of binary cross-entropy, specifically:

[0102] Where T is the total prediction step size; This is a pulse event indicator variable. It takes the value 1 when a pulse event occurs in time period t, and 0 otherwise.

[0103] The specific form of time regression loss is:

[0104] in, This represents the total number of pulse events that actually occurred in the current prediction batch; This means that the summation is only included when the indicator variable for the i-th pulse event is 1; This represents the actual event offset at step i; This refers to the time step added to the predicted time offset; This is the threshold parameter for Huber loss.

[0105] The specific form of amplitude prediction loss is as follows:

[0106] in, This represents the actual pulse amplitude at the i-th time step; The pulse amplitude prediction given by the amplitude prediction head; This is a smoothing term to prevent the logarithmic function from reaching negative infinity.

[0107] The specific forms of process constraint losses are:

[0108] in, This is a process requirement indicator function. It is 1 when the process state at time t violates the event triggering condition, and 0 otherwise.

[0109] The specific form of the cooperative scheduling constraint loss is as follows:

[0110] in, Total number of devices; Let be the probability that the nth device will generate a pulse during time period t; The maximum number of devices allowed to run simultaneously.

[0111] The consistency constraint loss is used to ensure that the outputs of the three prediction heads (event detection, time regression, and magnitude prediction) are consistent. For example, if the event probability is high, then the time offset and magnitude prediction should be reasonable (non-zero), specifically in the form of...

[0112] Training the above model can improve the accuracy of the agent in predicting impulse load.

[0113] Step 6: Based on the load forecast and production process status, determine the process adjustable state variables and critical pulse event locked variables, and evaluate the user's adjustability. After completing the basic load forecast and pulse load forecast, this embodiment further applies the dual-modal industrial load forecast results to the industrial load resource aggregation of the virtual power plant.

[0114] The virtual power plant aggregates industrial users with continuous base load and event-driven pulse load characteristics, such as steel, mining, machinery manufacturing, building materials processing, and chemical production.

[0115] For the m-th industrial user, its predicted total load in time period t can be expressed as:

[0116] in, This represents the predicted total load for the m-th industrial user during time period t. This represents the basic load forecast value; This represents the predicted value of the pulse load.

[0117] Furthermore, the aggregated industrial load forecast for the virtual power plant in time period t is as follows:

[0118] in, This represents the aggregated industrial load forecast of the virtual power plant during time period t; M represents the number of industrial users connected to the virtual power plant.

[0119] Because industrial loads are constrained by production processes and equipment status, virtual power plants cannot directly issue regulation commands based solely on predicted load magnitude. Therefore, this embodiment introduces the following process-adjustable state variables:

[0120] in, It is determined by the industrial user's production process parameters, equipment operating status, and event prediction results.

[0121] For pulsed loads, when predictions indicate the presence of an uninterruptible or unreducible critical pulse event during time period t, the virtual power plant should avoid issuing reduction commands to the core production equipment corresponding to that pulse event. Therefore, the critical pulse event locking variable is defined as follows:

[0122] Based on the above variables, the down-adjustment capability of the m-th industrial user in time period t is expressed as:

[0123] in, This represents the down-adjustment capability of the m-th industrial user during time period t; This represents the minimum allowable load that satisfies production safety, process continuity, and equipment operation constraints. This represents the maximum capacity reduction allowed for the m-th industrial user under contractual, equipment, or process constraints.

[0124] Accordingly, the upscaling capability of the m-th industrial user in time period t is expressed as:

[0125] in, This represents the upscaling capability of the m-th industrial user during time period t; Indicates the maximum allowable operating power of the equipment; This represents the maximum allowable capacity increase for the m-th industrial user.

[0126] Step 7: Aggregate the adjustable capabilities of each user to obtain the adjustable capacity boundary of the virtual power plant, and decompose the dispatch instructions based on the adjustable capacity boundary.

[0127] The adjustable capacity of the virtual power plant for aggregated industrial load in time period t is:

[0128] in, and These represent the aggregated industrial load up-adjustment capacity and aggregated down-adjustment capacity of the virtual power plant during time period t, respectively.

[0129] When a virtual power plant receives a regulation request from the grid dispatching agency or market operator, it aggregates the regulation request and distributes it to various industrial users. Let the virtual power plant's downward regulation request in time period t be... Then the downward adjustment instructions from each industrial user satisfy:

[0130] in, This indicates a downward adjustment instruction issued by the virtual power plant to the m-th industrial user.

[0131] Similarly, if the virtual power plant's upward adjustment demand in time period t is Then the upward adjustment instructions from each industrial user will satisfy:

[0132] in, This indicates an upward adjustment instruction issued by the virtual power plant to the m-th industrial user.

[0133] To prevent scheduling commands from disrupting critical production processes, when an industrial user is in a critical pulse event locked state, its corresponding downward scheduling command should satisfy the following:

[0134] Therefore, this invention can transform the bimodal industrial load forecast results into an industrial load aggregation forecast curve and regulation capacity boundary that can be directly used by the virtual power plant. This enables the virtual power plant to accurately grasp the available capacity of industrial load resources when participating in demand response, ancillary services, or electricity market transactions, and avoids impacting uninterrupted production processes.

[0135] Specific application examples: To further illustrate the application effect of the method in this embodiment in the scenario of industrial load resource aggregation in a virtual power plant, a park-level virtual power plant is used as an example. This virtual power plant connects to three types of industrial users: hot rolling mill users, mining hoisting users, and building material crushing users. Each industrial user has a dual-modal load characteristic with both base load and pulse load.

[0136] Among them, the basic load of hot rolling steel users mainly includes circulating water pumps, dust removal fans and air compressors, while the pulse load mainly comes from the main drive motor of hot rolling; the basic load of mining hoisting users mainly includes ventilation system and drainage system, while the pulse load mainly comes from the main shaft hoist; the basic load of building material crushing users mainly includes conveyor belt, dust removal system and auxiliary motor, while the pulse load mainly comes from the start-up of the crusher and the heavy-load crushing process.

[0137] Suppose that a virtual power plant performs load forecasting and regulation capacity assessment for the above three industrial users during a certain scheduling period t, with a forecasting time granularity of 15 minutes. Based on the basic load forecasting channel and pulse load forecasting channel of this invention, the forecast results for each industrial user during period t are shown in the table below:

[0138] As shown in the table above, hot-rolled steel users are predicted to experience steel biting or rolling pulse events during this period. Therefore, the key pulse event locking variables are:

[0139] There were no uninterrupted critical pulse events for mine hoisting users and building material crushing users during this period, therefore:

[0140]

[0141] Meanwhile, based on the production process status, equipment operating status, and user contract constraints, the process adjustability, minimum allowable load, maximum allowable load, and maximum adjustment capacity for each industrial user are shown in the table below:

[0142] The following calculations are performed for three industrial users based on the adjustable capacity calculation formula proposed in this embodiment.

[0143] For hot-rolled steel users:

[0144] This indicates that although there is room for reduction in load values ​​for hot-rolled steel users, the virtual power plant cannot treat its main drive pulse load as a resource that can be reduced due to the presence of critical pulse events during this period.

[0145] For users upgrading their mining operations:

[0146] For users of broken building materials:

[0147] Therefore, the aggregation downscaling capability of the virtual power plant in time period (t) is:

[0148] Furthermore, based on the adjustable capacity calculation formula proposed in this embodiment, the calculations for the three industrial users are as follows: For hot-rolled steel users:

[0149] For users upgrading their mining operations:

[0150] For users of broken building materials:

[0151] Therefore, the aggregation up-adjustment capacity of the virtual power plant in time period (t) is: 3.2

[0152] Meanwhile, the aggregated industrial load forecast for the virtual power plant during time period (t) is as follows:

[0153] The resulting aggregation of industrial load from the virtual power plant during this period is shown in the table below:

[0154] Furthermore, if the power grid dispatching agency issues peak-shaving requests to the virtual power plant during this period:

[0155] because:

[0156] Therefore, the virtual power plant has the capability to meet this peak shaving demand.

[0157] During the decomposition of scheduling instructions, the hot-rolled steel user is in a critical pulse event locked state:

[0158] Therefore, its downward adjustment instruction should satisfy:

[0159] Peak shaving demand is primarily borne by mining hoisting users and building material crushing users. The following allocation scheme can be adopted:

[0160]

[0161]

[0162] This allocation scheme satisfies: Simultaneously meet the adjustable capacity constraints of various industrial users:

[0163]

[0164]

[0165] Therefore, the decomposition result of the scheduling instruction meets the peak shaving requirements of the virtual power plant, while not affecting the critical pulse production events of the hot rolling steel users.

[0166] In another operational scenario, the power grid dispatching agency issues valley filling or load increase requests to the virtual power plant:

[0167] because:

[0168] Therefore, the virtual power plant has the capability to meet this increased load demand. The following upward adjustment command allocation scheme can be adopted:

[0169]

[0170]

[0171] This solution satisfies:

[0172] And satisfy:

[0173]

[0174]

[0175] As demonstrated by the above application examples, this embodiment can further calculate the adjustability of individual industrial users based on the dual-modal industrial load forecasting results, and aggregate the adjustability of multiple industrial users at the virtual power plant level. In peak shaving scenarios, this invention can identify critical pulse events occurring to hot-rolling steel users and set their adjustment capacity to zero, preventing the virtual power plant from erroneously reducing critical production loads. Simultaneously, by scheduling other industrial users in an adjustable state, the peak shaving requirements of the power grid can still be met. Compared to methods that estimate adjustment capacity solely based on total load size, this invention improves the accuracy, security, and executability of virtual power plant industrial load resource aggregation.

[0176] Example 2 Based on the method described in Embodiment 1, such as Figure 4 As shown, the purpose of this embodiment is to provide an industrial virtual power plant process event fusion prediction and regulation evaluation system, including: The data acquisition and decomposition module is used to collect multi-source data from industrial users and perform modal decomposition on the total load based on process event signals and load change characteristics, separating the basic load and the pulse load. The dual-modal load forecasting module is used to input the base load into a multi-scale temporal convolutional network and fuse it with a multi-attention gating mechanism to obtain the base load forecast value; input the pulse load into a multi-task learning framework to predict the probability of pulse events, the occurrence time offset, and the peak amplitude, and correct it with process constraints to obtain the pulse load forecast value; and obtain the total load forecast value by superimposing the base load forecast value and the pulse load forecast value. The adjustment capacity assessment and aggregation scheduling module is used to determine the process adjustable state variables and critical pulse event locked variables based on load forecast values ​​and production process status, assess the user's adjustability, aggregate the adjustability of each user, obtain the adjustable capacity boundary of the virtual power plant, and decompose the scheduling instructions according to the adjustable capacity boundary.

[0177] Example 3 The purpose of this embodiment is to provide a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the above-described method.

[0178] Example 4 The purpose of this embodiment is to provide a computer-readable storage medium.

[0179] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the steps of the above-described method.

[0180] The steps and methods involved in the apparatus of the above embodiments correspond to those in Embodiment 1. For specific implementation details, please refer to the relevant description section of Embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood as including any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and enabling the processor to perform any of the methods in this invention.

[0181] Those skilled in the art will understand that the modules or steps of the present invention described above can be implemented using general-purpose computer devices. Optionally, they can be implemented using computer-executable program code, thereby allowing them to be stored in a storage device for execution by a computer device, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. The present invention is not limited to any particular combination of hardware and software.

[0182] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for fusion prediction and regulation evaluation of process events in industrial virtual power plants, characterized in that, include: Collect multi-source data from industrial users, and perform modal decomposition on the total load based on process event signals and load change characteristics to separate the base load and pulse load; The baseline load is input into a multi-scale temporal convolutional network, and a multi-attention gating mechanism is fused to obtain the baseline load prediction value; The pulse load is input into a multi-task learning framework to predict the probability of pulse events, the occurrence time offset, and the peak amplitude. After correction by process constraints, the predicted value of the pulse load is obtained. The total load forecast is obtained by superimposing the base load forecast and the pulse load forecast. Based on load forecasts and production process status, adjustable state variables and key pulse event locked variables are determined, user adjustability is assessed, and the adjustability of each user is aggregated to obtain the adjustable capacity boundary of the virtual power plant. Dispatch instructions are then decomposed based on the adjustable capacity boundary.

2. The industrial virtual power plant process event fusion prediction and regulation evaluation method as described in claim 1, characterized in that, The collection of multi-source data from industrial users specifically includes: load power data, production process parameters, process event data, equipment status data, production plan data, and virtual power plant scheduling interface data.

3. The industrial virtual power plant process event fusion prediction and regulation evaluation method as described in claim 1, characterized in that, The modal decomposition of the total load includes: when the load increment at adjacent time moments exceeds a preset threshold and a corresponding process event signal exists at the same time, the load component is identified as a pulse load, and the pulse load is separated from the total load to obtain the base load.

4. The industrial virtual power plant process event fusion prediction and regulation evaluation method as described in claim 1, characterized in that, The multi-scale temporal convolutional network contains three parallel branches that capture short-period, medium-period, and long-period features, respectively; the multi-attention gating mechanism includes process attention, periodic attention, and event attention.

5. The industrial virtual power plant process event fusion prediction and regulation evaluation method as described in claim 1, characterized in that, The multi-task learning framework includes a shared feature encoder and three parallel task heads; The shared feature encoder employs a temporal convolutional network with asymmetric convolutional kernels and generates pulse feature context vectors by fusing event information with multi-head attention. The three task heads are an event detection head for predicting event probabilities, a time regression head for predicting occurrence time shifts, and an amplitude prediction head for predicting peak amplitudes.

6. The industrial virtual power plant process event fusion prediction and regulation evaluation method as described in claim 1, characterized in that, The critical pulse event locking variable is used to identify whether there are critical pulse events that cannot be interrupted or reduced within a time period.

7. The industrial virtual power plant process event fusion prediction and regulation evaluation method as described in claim 1, characterized in that, The assessment of user adjustability includes: defining process adjustability state variables, which are jointly determined by production process parameters, equipment operating status, and event prediction results, to indicate whether the process status allows load adjustment in the current period; defining critical pulse event lock-in variables to indicate whether there are any uninterrupted or unreducible critical pulse events in the current period; and calculating the user's down-adjustment capability based on the above variables. When the process status does not allow adjustment or there is a critical pulse event lock-in, the corresponding down-adjustment capability is zero.

8. An industrial virtual power plant process event fusion prediction and regulation evaluation system, characterized in that, include: The data acquisition and decomposition module is used to collect multi-source data from industrial users and perform modal decomposition on the total load based on process event signals and load change characteristics, separating the basic load and the pulse load. The dual-modal load forecasting module is used to input the base load into a multi-scale temporal convolutional network and fuse it with a multi-attention gating mechanism to obtain the base load forecast value. The pulse load is input into a multi-task learning framework to predict the probability of pulse events, the occurrence time offset, and the peak amplitude. After correction by process constraints, the predicted value of the pulse load is obtained. The total load forecast is obtained by superimposing the base load forecast and the pulse load forecast. The adjustment capacity assessment and aggregation scheduling module is used to determine the process adjustable state variables and critical pulse event locked variables based on load forecast values ​​and production process status, assess the user's adjustability, aggregate the adjustability of each user, obtain the adjustable capacity boundary of the virtual power plant, and decompose the scheduling instructions according to the adjustable capacity boundary.

9. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the industrial virtual power plant process event fusion prediction and regulation evaluation method according to any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it performs the steps of the industrial virtual power plant process event fusion prediction and regulation evaluation method as described in any one of claims 1-7.

Images