A virtual power plant power generation capacity prediction and optimization system based on timing feature enhancement

Abstract

The present application relates to the technical field of virtual power plant, in particular to a kind of virtual power plant power generation capacity prediction and optimization system based on timing feature enhancement.It includes multi-source data acquisition and screening unit, obtains virtual power plant whole network historical power and collects target controllable load operating state parameter;Timing feature enhancement unit is based on state parameter calculation system equivalent time lag parameter, filter reconstruction generates enhanced timing feature sequence;Dynamic prediction and compensation unit inputs network with enhanced sequence and obtains initial prediction, uses the power deviation rate of physical execution to carry out numerical compensation and obtains target output prediction sequence;Optimization scheduling unit constructs multi-objective optimization model to generate instruction with cost minimization;Closed-loop correction unit issues instruction and back sampling actual output to update deviation rate.By time lag feature enhancement and feedforward compensation to the mechanical and thermal dynamic inertia of equipment, the distortion misplacement problem of traditional prediction in power mutation interval is eliminated.

Description

Technical Field

[0001] This invention relates to the field of virtual power plant technology, and more specifically, to a virtual power plant power generation capacity prediction and optimization system based on time-series feature enhancement. Background Technology

[0002] With the increasing proportion of new energy sources in the new power system, virtual power plants (VPPs), as an important technical means to aggregate distributed energy resources, energy storage, and controllable loads to participate in grid dispatch, have received widespread attention and engineering applications. However, existing virtual power plant prediction and dispatch systems still face the following significant technical bottlenecks in actual industrial-grade control scenarios: Existing virtual power plant cloud platforms typically treat underlying equipment as ideal nodes capable of responding to commands instantly. However, in real-world industrial scenarios, heavy loads such as coal conveyor belts in mines and large industrial electric boilers, constrained by their enormous mechanical inertia or thermal energy storage, cannot achieve "rapid power reduction in a short period." When the power grid issues a drastic load reduction command, if the scheduling algorithm fails to factor in the equipment's "physical sluggishness and braking inertia" beforehand, it will lead to a severe disconnect between the virtual power plant's actual grid connection point power calculation curve and the power grid's assessment curve. This can easily result in hefty penalties for breach of contract and even the risk of being expelled from the spot and ancillary services market. Meanwhile, to improve prediction accuracy, existing systems extensively employ purely data-driven neural network models. However, in the face of sudden changes in grid demand or drastic shifts in operating conditions, black-box models often provide extreme value predictions that defy physical principles. If such dispatch commands containing distorted power slopes are issued directly without verification, the underlying control system in the field will forcefully follow the commands, leading to severe power oscillations in motor output. This can easily trigger the inverter's overcurrent protection and cause it to trip, seriously threatening the continuous production safety of the industrial site. Therefore, this paper proposes a virtual power plant power generation capacity prediction and optimization system based on time-series feature enhancement. Summary of the Invention

[0003] The purpose of this invention is to provide a virtual power plant power generation capacity prediction and optimization system based on time-series feature enhancement, in order to solve the problem mentioned in the background art that the underlying control system, in order to forcibly track the command, will cause the motor output to oscillate violently, which will easily trigger the inverter overcurrent protection and trip, seriously threatening the continuous production safety of the industrial site.

[0004] To achieve the above objectives, the present invention aims to provide a virtual power plant generation capacity prediction and optimization system based on time-series feature enhancement, comprising: A multi-source data acquisition and filtering unit is used to acquire the historical power sequence of the entire network of the virtual power plant in real time, and to filter out the target controllable load from the virtual power plant according to the preset capacity ratio threshold, and to collect the real-time operating status parameters of the target controllable load. A time-series feature enhancement unit calculates the system equivalent time delay parameters of the target controllable load based on real-time operating status parameters, filters and reconstructs the historical power sequence of the entire network, and generates an enhanced time-series feature sequence. The dynamic prediction and compensation unit is used to input the enhanced time series feature sequence into a pre-trained time series prediction network and output an initial output prediction sequence; and by extracting the actual executed power and predicted power of the virtual power plant in the previous scheduling cycle, calculating the power deviation change rate between the two, and using the power deviation change rate to numerically compensate the initial output prediction sequence to obtain the target output prediction sequence. An optimized scheduling unit calculates the available adjustment margin of the virtual power plant in the future scheduling period based on the target output prediction sequence. It constructs a multi-objective optimization model with the equipment operating boundary conditions and available adjustment margin as constraints and the goal of minimizing scheduling operation costs, and solves the model to generate optimized scheduling instructions. A closed-loop correction unit is used to issue and execute the optimized scheduling command, monitor the actual output response data of the virtual power plant in real time, and update the power deviation change rate based on the actual output response data.

[0005] As a further improvement to this technical solution, the target controllable load is divided into heavy machinery rotation load and temperature control and heat processing load according to its power response time delay mechanism; Wherein, when the target controllable load is the heavy machinery rotation load, the real-time operating status parameters include at least the actual speed of the motor rotor, the stator side operating current, and the mechanical shaft end load torque; When the target controllable load is the temperature control and heat processing type load, the real-time operating status parameters include at least the current input power, the temperature difference between the supply and return water circuits of the circulating medium, and the deviation between the actual temperature of the target process environment and the set temperature.

[0006] As a further improvement to this technical solution, the time-series feature enhancement unit includes a parameter calculation and aggregation module and a filtering and reconstruction module; The parameter calculation and aggregation module calculates the load response time constant of the target controllable load based on real-time operating status parameters, and calculates the equivalent time delay parameters of the system by weighting the load response time constant and the current actual power ratio of the target controllable load. The filtering and reconstruction module is used to convert the system's equivalent time delay parameters into first-order inertial filter coefficients, and to perform iterative smoothing processing on each time-series sampling node of the entire network's historical power sequence using a discrete differential filtering algorithm, outputting an enhanced time-series feature sequence.

[0007] As a further improvement to this technical solution, the specific steps involved in calculating the load response time constant of the target controllable load are as follows: When the target controllable load is the heavy machinery rotation load, the load response time constant is calculated based on the dynamic ratio of the actual rotational speed of the motor rotor to the load torque at the mechanical shaft end, combined with the preset equipment rotational inertia equivalent; wherein, the load response time constant is positively correlated with the actual rotational inertia of the motor rotor and negatively correlated with the magnitude of the resultant force of the load torque at the mechanical shaft end and the driving electromagnetic torque. When the target controllable load is a temperature control and heat processing load, the corresponding load response time constant is determined based on the deviation between the actual temperature and the set temperature of the target process environment, the equivalent temperature difference ladder formed by the temperature difference of the supply and return water circuits of the circulating medium, and the preset equivalent heat capacity of the system and the current input active power of the core working element.

[0008] As a further improvement to this technical solution, the calculation process of the enhanced time-series feature sequence is as follows: Collect the current actual power consumption of each target controllable load and calculate its actual power proportion in the original total power of the virtual power plant network. The actual power ratio is used as a dynamic weighting coefficient to perform a weighted summation of the calculated load response time constants. By integrating the pre-calibrated reference time constants of the remaining background devices in the entire network, the equivalent time delay parameters of the system are calculated; Using the system's equivalent time delay parameters as the smoothing constraint benchmark for the first-order inertial filtering algorithm, dynamic filtering coefficients are constructed by combining the system data sampling period. By using dynamic filtering coefficients to perform sliding iterative reconstruction on each discrete sampling point of the historical power sequence of the entire network, an enhanced time series feature sequence is finally obtained.

[0009] As a further improvement to this technical solution, the dynamic prediction and compensation unit includes an initial prediction module and a dynamic compensation module; The initial prediction module is used to input the enhanced temporal feature sequence into a pre-trained long short-term memory temporal prediction network, and to perform sequence mapping and feature decoding of the forward time step using a gating mechanism, and output the initial output prediction sequence. The dynamic compensation module is used to extract the actual power sequence and the historical predicted power sequence of the virtual power plant in the previous scheduling cycle, and to calculate the power deviation change rate between the two using the discrete difference algorithm. Simultaneously, a sliding window threshold detection algorithm is used to identify power mutation nodes in the initial power output prediction sequence, and the power deviation change rate is used as a feedforward physical correction term to perform reverse cancellation and numerical compensation on the predicted values ​​at the power mutation nodes in the initial power output prediction sequence, and finally output the target power output prediction sequence.

[0010] As a further improvement to this technical solution, the specific calculation logic for obtaining the target output prediction sequence through numerical compensation in the dynamic compensation module is as follows: The difference between the actual execution power sequence and the historical predicted power sequence is calculated at discrete time steps to obtain the dynamic error sequence; The power deviation change rate is calculated by performing a discrete first-order difference operation on the dynamic error sequence. The initial output prediction sequence is traversed using a sliding observation window to calculate the slope of the power difference between adjacent prediction time steps; When the absolute value of the power difference slope is greater than the preset physical ramp rate threshold, the corresponding prediction time step is marked as a power mutation node. At the power mutation node, a preset closed-loop feedback gain coefficient is introduced to convert the power deviation rate of change into a reverse compensation power term. The initial predicted power value corresponding to the power mutation node is compensated by reverse power compensation, and the compensated sequence is smoothly reconstructed to finally obtain the target output prediction sequence.

[0011] As a further improvement to this technical solution, the optimized scheduling unit includes a margin calculation module and a multi-objective optimization module; The margin calculation module is used to receive the target output prediction sequence and simultaneously obtain the real-time equipment operation boundary conditions of each target controllable load inside the virtual power plant. The target output prediction sequence is used as the benchmark output evolution trajectory and is physically superimposed and optimized spatial intersection calculated with the equipment operation boundary conditions to calculate the available adjustment margin of the virtual power plant in the future scheduling period. The available adjustment margin includes the upward adjustment power adjustment margin and the downward adjustment power adjustment margin. Among them, the equipment operation boundary conditions include at least the absolute power amplitude boundary, the dynamic ramp rate boundary, the energy state and process tolerance boundary, and the start-stop cycle and service life boundary. The multi-objective optimization module receives the available adjustment margin and uses it as the underlying hard safety constraint. It constructs a multi-objective optimization model with the objective function of minimizing the overall scheduling and operation cost of the virtual power plant. The multi-objective optimization model is solved by a mathematical programming algorithm to calculate the optimal power allocation strategy. The strategy is then converted into optimized scheduling instructions for each controlled device in the virtual power plant and issued.

[0012] As a further improvement to this technical solution, the closed-loop correction unit updates the power deviation change rate, involving the following steps: Extract the optimization scheduling instruction of the first future time step output by the multi-objective optimization module, parse its protocol and convert it into a control message that conforms to the communication standard of the underlying controlled device, and drive the corresponding physical device to perform the actual power adjustment action; By deploying intelligent acquisition terminals at the bottom layer, the actual power data of the virtual power plant in the current scheduling cycle is collected in real time at a preset sampling frequency, and aggregated to form an actual output response data sequence; The actual output response data sequence is fed back to the dynamic prediction and compensation unit; the dynamic prediction and compensation unit extracts the actual execution power observation value in the previous scheduling cycle, uses it as the historical known condition for the latest scheduling cycle, and calculates the actual execution power observation value and the corresponding historical predicted power value to obtain a dynamic error sequence used to characterize the physical execution deviation. Based on the dynamic error sequence, the dynamic prediction and compensation unit is driven to continuously update the average power deviation change rate, thereby realizing feedforward physical correction and global closed-loop self-healing of the initial output prediction sequence in the next prediction window.

[0013] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. In this virtual power plant power generation capacity prediction and optimization system based on time-series feature enhancement, a time-series feature enhancement unit is introduced. By extracting the actual operating status of the underlying equipment (such as heavy machinery and thermal power equipment), the equivalent time delay parameters of the system are dynamically calculated and reconstructed by discrete first-order inertial filtering. This mechanism enables the model to implicitly integrate the physical sluggishness and motion inertia of the equipment at the beginning of the prediction, effectively avoiding the issuance of "empty promises" by the dispatch cloud. It solves the problem of the actual grid connection curve being out of sync with the grid assessment curve due to the "slow response and inability to stop" of the on-site equipment, ensuring a high performance rate of the virtual power plant.

[0014] 2. In this virtual power plant power generation capacity prediction and optimization system based on time-series feature enhancement, addressing the pain point of pure black-box neural networks being prone to distortion extremes in command mutation intervals, this invention constructs a composite architecture of "AI decoding + physical ramp threshold identification + closed-loop error compensation" in the dynamic prediction and compensation unit. By identifying mutation nodes that violate the ramp rate limits of on-site motors or processes, and using the real error evolution trend of the previous cycle for feedforward directional cancellation, this system eliminates safety accidents such as severe power oscillations in underlying equipment and inverter tripping caused by blindly issuing extreme power ramp rate commands, significantly ensuring continuous production safety in industrial sites.

[0015] 3. In this virtual power plant power generation capacity prediction and optimization system based on time-series feature enhancement, when constructing a multi-objective optimization scheduling model, the absolute power change of the equipment and the depreciation penalty coefficient based on the underlying wear characteristics are coupled into the objective function, transforming the implicit physical wear of the equipment into a quantifiable economic constraint; giving the solver the business intelligence of tiered scheduling, automatically avoiding the frequent issuance of cross-condition start-up and shutdown or deep charging and discharging commands for energy storage for small electricity price differences, thereby significantly extending the service life of the core heavy assets inside the virtual power plant and maximizing the comprehensive economic benefits throughout the entire life cycle. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the overall process of the present invention.

[0017] The meanings of the various markings in the diagram are as follows: 1. Multi-source data acquisition and filtering unit; 2. Temporal feature enhancement unit; 21. Parameter calculation and aggregation module; 22. Filtering and reconstruction module; 3. Dynamic prediction and compensation unit; 31. Initial prediction module; 32. Dynamic compensation module; 4. Optimized scheduling unit; 41. Margin calculation module; 42. Multi-objective optimization module; 5. Closed-loop correction unit. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0019] Please see Figure 1 As shown, a virtual power plant generation capacity prediction and optimization system based on time-series feature enhancement is provided, including: Multi-source data acquisition and filtering unit 1 is used to acquire the network-wide historical power sequence of the virtual power plant in real time (the network-wide historical power sequence is obtained by acquiring active and reactive power data within the past 24 hours or longer preset historical time window at a preset sampling frequency (e.g., 1 minute or 5 minutes as a sampling period) to form a time sequence; further, after acquiring the sequence, the system uses conventional algorithms such as linear interpolation or mean filling to preprocess outliers or missing values ​​in the sequence to ensure the numerical stability of subsequent first-order inertial filter reconstruction), and selects target controllable loads from the virtual power plant according to a preset capacity ratio threshold, and collects the real-time operating status parameters of the target controllable loads; The target controllable load is an electrical device that exhibits power response time delay characteristics due to physical mechanical inertia or thermodynamic inertia. It is worth noting that, in specific embodiments of this application, an electrical device that exhibits power response time delay characteristics due to physical mechanical inertia or thermodynamic inertia refers to a device that, upon receiving a power adjustment command (such as a frequency reduction, shutdown, or power increase command) from a virtual power plant, cannot instantly jump to the target power value due to the energy storage effect or motion inertia of its internal physical structure. Instead, it exhibits a gradual transition process with a significant time constant (e.g., a response time greater than a preset 10-15 seconds). Specifically, to enable those skilled in the art to clearly understand and implement this invention, the target controllable load is divided into the following two specific clusters of electrical devices based on the physical mechanism of its time delay: Category 1: Heavy-duty rotating equipment limited by physical mechanical inertia (the time-delay characteristic of this type of equipment stems from its huge mechanical rotor rotational inertia and the mechanical inertia of the driven load. When the frequency converter issues a frequency reduction command, due to the conservation of kinetic energy, its rotational speed and actual electrical power exhibit first-order or multi-order inertial decay; in industrial supply chain scenarios (such as energy extraction, washing and processing, and bulk logistics processing), this type of equipment includes, but is not limited to: Main ventilation unit: It has a large-mass impeller. When adjusting the air volume to reduce power, the flywheel effect of the impeller will cause the power curve to slowly decrease. Long-distance continuous material conveyors (such as main coal flow belt conveyors): The belt carries tens to hundreds of tons of bulk materials (such as coal and ore). When decelerating or participating in peak shaving response, the braking and power reduction process must be physically lengthened to prevent material accumulation or belt tearing. Large fluid transport pumping stations (such as heavy medium washing pumps and mine main drainage pumps): Due to the water hammer effect and fluid resistance of the pipeline network, there is a significant mechanical response delay in the acceleration and deceleration process of their motors. The second category: Temperature control and heat processing equipment limited by thermodynamic inertia (the time delay characteristic of this type of equipment stems from the thermal energy storage buffering effect generated by the large specific heat capacity of the working medium (such as water, heat transfer oil, air). When the power is cut off or the heating / cooling power is reduced, the heat / cold energy accumulated inside the system can still maintain the process requirements for a period of time, and the power adjustment will not cause a sudden change in the environmental state, exhibiting a power response delay of several minutes to tens of minutes), specifically including but not limited to: Bulk material drying systems (such as coal slime dryers and grain drying towers): their heating furnaces or drying drums have extremely high thermal inertia, and their power regulation when participating in virtual power plant scheduling exhibits typical thermodynamic hysteresis. Central heating / electric heating boilers: rely on a large water circulation system for heat transfer, and the high specific heat capacity of water makes it a natural thermal energy storage carrier; Deep cooling or large-scale refrigeration units: These include compressors, condensers, and a vast chilled water piping network. The phase change process of the refrigerant and the residual cooling of the piping network give them excellent power regulation flexibility and time-delay characteristics.

[0020] Furthermore, to avoid including all small inertial devices (such as small lighting fans) in the highly complex time delay parameter calculation, this system sets a capacity threshold for the target controllable load. In actual engineering deployment, by reading the equipment ledger, only large equipment whose rated power accounts for a proportion greater than the preset capacity proportion threshold (e.g., 5%~10%) of the total capacity of the virtual power plant is selected as the target controllable load. This ensures that the equivalent time delay parameter can truly reflect the dominant inertia of the entire network, and greatly reduces the data communication pressure and edge computing load of the system.

[0021] Furthermore, the target controllable load is divided into heavy machinery rotation load and temperature control and heat processing load according to its power response time delay mechanism (i.e., heavy machinery rotation equipment limited by physical mechanical inertia, and temperature control and heat processing equipment limited by thermodynamic inertia). Wherein, when the target controllable load is the heavy machinery rotation load, the real-time operating status parameters include at least the actual speed of the motor rotor, the stator side operating current, and the mechanical shaft end load torque; When the target controllable load is the temperature control and heat processing type load, the real-time operating status parameters include at least the current input power, the temperature difference between the supply and return water circuits of the circulating medium, and the deviation between the actual temperature of the target process environment and the set temperature.

[0022] In this embodiment, the virtual power plant generation capacity prediction and optimization system based on time-series feature enhancement also includes a time-series feature enhancement unit 2. The time-series feature enhancement unit 2 calculates the system equivalent time delay parameters of the target controllable load based on real-time operating status parameters, and uses the system equivalent time delay parameters to filter and reconstruct the historical power sequence of the entire network to generate an enhanced time-series feature sequence that eliminates physical execution delay misalignment. The temporal feature enhancement unit 2 includes a parameter calculation and aggregation module 21 and a filtering and reconstruction module 22. Among them, the parameter calculation and aggregation module 21 calculates the load response time constant of the target controllable load based on the real-time operating status parameters, and calculates the system equivalent time delay parameter characterizing the overall physical lag of the virtual power plant based on the load response time constant and the current actual power ratio of the target controllable load. The filtering and reconstruction module 22 is used to convert the system equivalent time delay parameters into first-order inertial filter coefficients, and to use the discrete differential filtering algorithm to iteratively smooth each time-series sampling node of the historical power sequence of the entire network, and output an enhanced time-series feature sequence that eliminates physical time delay misalignment. It is worth noting that, in this embodiment, the specific steps involved in calculating the load response time constant of the target controllable load are as follows: When the target controllable load is the heavy machinery rotation load, the corresponding load response time constant is determined based on the dynamic ratio of the actual rotational speed of the motor rotor to the load torque at the mechanical shaft end, combined with the preset equipment rotational inertia equivalent; wherein, the load response time constant is positively correlated with the actual rotational inertia of the motor rotor and negatively correlated with the magnitude of the resultant force of the load torque at the mechanical shaft end and the driving electromagnetic torque. Specifically, in this embodiment, when the first When the target controllable load is heavy mechanical rotating equipment, its load response time constant is calculated based on the dynamic equation:

[0023] in, Indicates the first The preset equivalent of the rotational inertia of the target controllable load is used to characterize the inherent limiting property of the equipment in maintaining its current rotational kinetic energy, and its dimension is kilogram-square meter (kg / m²). Specifically, the preset value is based on the equivalent conversion of the motor nameplate parameters of the target controllable load combined with the mechanical transmission ratio, or it is an objective physical quantity directly measured by a person skilled in the art, which is obtained and entered in advance during the system initialization stage through the field no-load identification program of the industrial frequency converter. The typical range of the preset equipment rotational inertia equivalent is usually between 50 kg·m² and 5000 kg·m². The angular velocity corresponding to the actual rotational speed of the motor rotor, expressed in radians per second ( ). ); for The actual operating current on the stator side of the motor at any given time is expressed in amperes (A). ); This is the electromagnetic torque conversion proportionality coefficient, used to linearly map the motor stator current to the electromagnetic drive torque, with dimensions in Newton-meters per ampere (Nm / ampere). ); for The real-time load torque at the end of the mechanical shaft represents the magnitude of the mechanical load on which the equipment does work against external resistance, and its dimension is Newton-meter (Nm). ); For the preset minimum positive constant (usually taken as...) This is used to ensure the numerical stability of the algorithm and prevent overflow errors when the system is in a steady state (driving force and resistance are balanced, and the denominator approaches zero). express Time of the first The load response time constant of heavy rotating machinery is used to characterize the physical hysteresis time required for the equipment to change its current state of motion, and its dimension is seconds. ).

[0024] When the target controllable load is a temperature control and heat processing load, the corresponding load response time constant is determined based on the deviation between the actual temperature and the set temperature of the target process environment, the equivalent temperature difference ladder formed by the temperature difference of the supply and return water circuits of the circulating medium, and the preset equivalent heat capacity of the system and the current input active power of the core working element.

[0025] In another embodiment, when the first When the target controllable load is a temperature-controlled or heat-processing type load, its load response time constant is calculated based on the heat transfer equation as follows:

[0026] in, For the first The preset system equivalent heat capacity of the target controllable load is used to characterize the heat storage and buffering capacity of the device and its internal circulating medium, and its dimension is joules per Kelvin (J / K). The preset system equivalent heat capacity is based on theoretical calculations of static material properties or data identification based on dynamic step response. Specifically, it extracts the mass and corresponding atmospheric pressure specific heat capacity parameters of each heat storage component (such as metal pipe wall metal and internal fluid medium) within the target controllable load. The total system heat capacity is calculated using a multi-medium heat capacity equivalent aggregation algorithm (or mass weighted summation algorithm). Alternatively, it obtains the natural heat dissipation temperature time series data of the equipment after steady-state operation and sudden power failure, fits its natural cooling curve using a first-order heat conduction model (such as Newton's law of cooling), calculates the temperature decay time constant under the current physical state, and then combines the equipment's calibrated thermal resistance to calculate the dynamic equivalent heat capacity value that conforms to the equipment's aging status. The typical range of the preset system equivalent heat capacity is usually within 10. 6 J / K to 10 8 The J / K ranges from megajoules per kelvin to hundreds of megajoules per kelvin. for The absolute value of the deviation between the actual temperature of the target process environment and the set temperature at any given time, expressed in Kelvin (K). ); This represents a dimensionless temperature difference weighting correction coefficient, used to balance the influence of ambient temperature deviation and pipeline circulation temperature difference on the overall thermal inertia calculation of the system. for The temperature difference in the supply and return water circuits of the circulating medium (such as the temperature difference between supply and return water, and between inlet and outlet air) is measured in Kelvin. ); for The current active power input of the core working element (such as the heating element of an electric boiler or a refrigeration compressor), measured in watts (W). ); This is used to prevent extremely small positive numbers with a denominator of zero; express Time of the first The load response time constant for temperature control and heat processing loads, in seconds ( ).

[0027] In this embodiment, after performing corresponding calculations based on the physical type of the equipment, a uniform value is assigned to the load response time constant of that node. , This is used for subsequent calculations of the system's equivalent time delay parameters.

[0028] Furthermore, the specific calculation process for enhancing time-series feature sequences is as follows: Collect the current actual power consumption of each target controllable load and calculate its actual power proportion in the original total power of the virtual power plant network. The actual power ratio is used as a dynamic weighting coefficient to perform a weighted summation of the calculated load response time constants. By integrating the pre-calibrated reference time constants of the remaining background devices in the entire network, the equivalent time delay parameters of the system, which characterize the macroscopic sluggishness of the entire network, are calculated. Using the system's equivalent time delay parameters as the smoothing constraint benchmark for the first-order inertial filtering algorithm, dynamic filtering coefficients are constructed by combining the system data sampling period. By using dynamic filtering coefficients to perform sliding iterative reconstruction on each discrete sampling point of the historical power sequence of the entire network, an enhanced time series feature sequence is finally obtained.

[0029] In this embodiment, the load response time constant of each of the target controllable loads is calculated. The equivalent time delay parameters of the system at time t. : ; in, express The system equivalent time delay parameter, which characterizes the overall macroscopic physical lag of the virtual power plant, is measured in seconds. ); This represents the total number of target controllable loads selected from the entire network by the virtual power plant, and is a positive integer. The index number represents the target controllable load. ; for The virtual power plant corresponding to the original total power of the entire network at the specified time point in the historical power sequence of the entire network; express Time of the first The actual power consumption of a target controllable load, in kilowatts (kW) ); The reference time constant for the remaining background equipment in the entire network, pre-calibrated, is used to characterize the average response time constant of the remaining background load (such as basic lighting, light office equipment, etc.) of the virtual power plant after removing all screened key target controllable loads. The unit is seconds. Specifically, its typical value ranges from 0.1 seconds to 5 seconds.

[0030] Furthermore, the system's data sampling period is set to... Then The equivalent time delay parameters of the system calculated at each time step , construct corresponding Discretized first-order inertial dynamic filter coefficients at time points : ; extract The original power sample value corresponding to the historical power sequence of the entire network at the specified time. Calculated using a first-order discrete difference filtering algorithm Enhanced temporal feature sequence at time step : ; in, for The reconstructed numerical values ​​of the enhanced temporal feature sequence at time step; express The discretized first-order inertial dynamic filter coefficients constructed by the time-matter system are dimensionless parameters because the numerator and denominator both have the dimension of time, and their values ​​always satisfy the following: ; This indicates the data sampling period set by the system (i.e., the time interval step between two adjacent data acquisitions). In this embodiment, when the virtual power plant participates in the fast-response ancillary service market such as AGC (Automatic Generation Control) or primary frequency regulation of the power grid, the... The typical value range is set as follows (seconds); When the virtual power plant participates in the spot electricity market and performs routine peak shaving and valley filling or load tracking tasks, in order to reduce unnecessary communication overhead and database storage pressure, the system collects data through conventional SCADA or smart meters. The typical value range is broadened and set to (Right now The unit of measurement is seconds. ); express The enhanced timing characteristic sequence value output after filtering and reconstruction by this system is the smoothed power value after eliminating physical time delay misalignment; By performing sliding iterative filtering on the historical power sequence of the entire network within the historical time window, the complete enhanced time series feature sequence is output, so that the abrupt slope of the reconstructed time series feature is strictly limited by the equivalent time delay parameter of the system.

[0031] The virtual power plant power generation capacity prediction and optimization system based on time-series feature enhancement also includes a dynamic prediction and compensation unit 3. The dynamic prediction and compensation unit 3 is used to input the enhanced time-series feature sequence into a pre-trained time-series prediction network and output an initial output prediction sequence. By extracting the actual and predicted power of the virtual power plant in the previous scheduling cycle, calculating the power deviation rate between the two, and using the power deviation rate to numerically compensate the power mutation nodes in the initial output prediction sequence, the target output prediction sequence is obtained. In this embodiment, the dynamic prediction and compensation unit 3 includes an initial prediction module 31 and a dynamic compensation module 32; The initial prediction module 31 is used to input the enhanced temporal feature sequence into the pre-trained long short-term memory temporal prediction network, and to perform sequence mapping and feature decoding of the forward time step using a gating mechanism, and output the initial output prediction sequence. In this embodiment, the pre-training steps of the pre-trained Long Short-Term Memory Temporal Prediction Network in the initial prediction module 31 are as follows: Massive, multi-dimensional operation logs of virtual power plants during typical historical operating cycles are extracted. Historical feature data is input into a pre-built physical enhancement module for time-series features, generating enhanced historical time-series feature sequences with physical time delays as input feature samples for the model. Simultaneously, the actual historical power output sequences of the virtual power plant's grid-connected points within the corresponding time periods are extracted as the model's target true labels. A sliding data window technique is used to segment the aforementioned time-series sequences, dividing them into training and validation sets according to a preset ratio (e.g., typically 80% for the training set and 20% for the validation set). The topology of the Long Short-Term Memory (LSTM) temporal prediction network is initialized on a cloud server. Its internal nodes are configured with Sigmoid-gated activation functions and Tanh state activation functions for nonlinear mapping, and random initial weight matrices and bias parameters are assigned to the forget gate, input gate, output gate, and linear mapping layer. The historical enhanced temporal feature sequences from the training set are input into the network in batches according to time steps. Through gated evolution of cell states, the corresponding historical prediction sequences are output forward. The mean squared error (MSE) or mean absolute error (MAE) is selected as the target loss function to calculate the temporal reconstruction error between the historical prediction sequence currently output by the network and the target true label. Subsequently, a time-based backpropagation algorithm is used to calculate the gradient vector of this error sequence with respect to the weights and biases of each gate structure within the network. An adaptive gradient descent optimization algorithm (such as the Adam optimizer) is introduced. Based on the calculated gradient vector, all weight matrices and bias parameters within the network are updated in reverse with a preset learning rate (set to the order of 0.001 to 0.01). After each training round, a validation set is input into the network to calculate the generalization error. When the validation set error no longer decreases significantly with the number of iterations and tends to stabilize, or when the maximum preset number of iterations is reached, an early stopping mechanism is triggered to end the training. Finally, all node parameters within the network at this point are solidified, resulting in the pre-trained Long Short-Term Memory temporal prediction network, which is then deployed to the temporal decoding and initial prediction submodule 31 for online invocation.

[0032] Furthermore, in this embodiment, the pre-trained Long Short-Term Memory (LSTM) temporal prediction network is invoked: The enhanced temporal feature sequence within the preset historical observation window before the current prediction time is extracted, and it is divided into temporal input feature tensors according to time steps. The initial hidden state and initial cell state of the network are set to zero. The enhanced temporal feature sequence has already incorporated the physical response time constants of each controllable load at the bottom layer. The data received by the network not only includes power amplitude fluctuations, but also implicitly carries the physical time delay mechanism of the device. First, the hidden state of the previous time step and the input features of the current time step are read through the internal forget gate logic. The effectiveness of the historical long-term memory is adaptively evaluated through a nonlinear activation function. Then, the retention weight is calculated, and historical redundant features that are no longer related to the current power evolution trend (such as small perturbations caused by old scheduling instructions that have ended) are physically truncated and removed. After removing invalid history, the enhanced temporal feature sequence input at the current time is evaluated through input gate logic; through the combined action of dual activation functions (Sigmoid and Tanh dual activation functions), on the one hand, the core time delay features in the current input data that have a decisive influence on the future power trend are identified, and on the other hand, corresponding candidate state increments are generated to determine how much fresh physical mechanism information needs to be formally recorded into the network's long-term memory. The core cell state of this network (i.e., the long-term memory stream) is transmitted through a linear channel that spans the entire time step. The historical valid state retained by the forget gate is superimposed and fused with the current state increment extracted by the input gate to complete the rolling update of the cell state. This information flow process is equivalent to and simulates, at the physical level, the accumulation of thermal energy in the thermodynamic medium inside a virtual power plant or the gradual accumulation of kinetic energy in heavy machinery. Finally, the network calculates the hidden state output for the current time step based on the updated core cell state and the current input features through the output gate logic; it then extracts the final hidden state vector from the last time step of the historical observation window and directly feeds it into the subsequent fully connected output layer. The dimension of the output neuron in this fully connected layer is strictly set to the total step size of the future prediction window. This allows for the simultaneous mapping of highly condensed hidden state features into a single, parallel mapping that includes... The future power values ​​at each time point are arranged in order of time step to form an unfiltered initial power prediction sequence, which is then transmitted to the downstream dynamic compensation module 32. In this embodiment, the dynamic compensation module 32 is used to extract the actual executed power sequence and the historical predicted power sequence of the virtual power plant in the previous scheduling cycle, and to calculate the power deviation change rate between the two using the discrete difference algorithm. Simultaneously, a sliding window threshold detection algorithm is used to identify power mutation nodes in the initial power output prediction sequence, and the power deviation change rate is used as a feedforward physical correction term to perform reverse cancellation and numerical compensation on the predicted values ​​at the power mutation nodes in the initial power output prediction sequence, and finally outputs the target power output prediction sequence to eliminate dynamic following error.

[0033] Furthermore, in the dynamic compensation module 32, the specific calculation logic for obtaining the target output prediction sequence through numerical compensation is as follows: Set the current prediction task trigger time as The time steps for discrete data sampling and prediction are both Then extract the time observation window corresponding to the previous scheduling cycle. Inside (of which) The actual executed power sequence and historical predicted power sequence of the virtual power plant (the total number of historical time steps included in the scheduling period) are calculated by discrete time steps. The difference between the two is calculated point by point, and the dynamic error sequence is obtained. : ; in, ; In the formula, For the first Actual execution power observations for each step; For the system in the previous cycle targeting the first The historical predicted values ​​output step by step; Indicates the first The dynamic power deviation value of the step; when Time indicates that the model prediction is lagging, when The time indicates that the model prediction has overshooted; To determine the divergence or convergence trend of the quantization error, the dynamic error sequence is... Discrete first-order difference operations are performed; specifically, to avoid high-frequency single-point noise interference, the dynamic error sequence is truncated. Approaching the current moment The average power deviation rate of change in the quantitative characterization model's prediction of overshoot or hysteresis divergence was calculated from the end data window. As a macroscopic deviation characteristic: ; In the formula, The preset tail smoothing window length (the value satisfies) ), This indicates the total number of historical time steps included in the scheduling period; Index variable representing the time step; Indicates the time elapsed since the current time Reverse calculation Dynamic power deviation sampling points at each time step. Indicates the time elapsed since the current time Reverse calculation Dynamic power deviation sampling points at each time step. and All are from dynamic error sequences The adjacent discrete error sampling points are extracted in reverse order. It represents the rate of recent predicted overshoot or hysteresis divergence, and its dimension is power / time (e.g., MW / s). The initial output prediction sequence is traversed using a sliding observation window to calculate the slope of the power difference between adjacent prediction time steps; When the absolute value of the power difference slope is greater than the preset physical ramp rate threshold, the corresponding prediction time step is marked as a power mutation node. In this embodiment, the future initial output prediction sequence is assumed to be: ,in For future prediction step size sequence, The total prediction window length (i.e., the total number of future prediction steps) is used; a sliding window is used to traverse the sequence, and the slope of the power difference between adjacent prediction time steps is calculated. : ; In the formula, Indicates future time The slope of the power difference between adjacent prediction time steps; Indicates the future number The initial output prediction value for each prediction time step; Indicates the future number The initial output prediction value for each prediction time step; This represents the time step for discrete data sampling and prediction. In specific engineering implementations, for industrial virtual power plants whose main loads are heavy-duty fans, water pumps, or centralized electric boilers, the power changes under normal operating conditions rarely exhibit a step-like state; therefore, the physical ramp rate threshold... The typical empirical value range is between 0.1MW / s and 1.5MW / s (based on the nameplate parameters of each controlled device in the virtual power plant and the maximum physical ramping capability of the unit, and the aggregated weighted calibration). Predicted values ​​exceeding this slope are forcibly identified as numerical anomaly nodes in the AI ​​model. The calculated absolute value of the slope Compared with the physical ramp rate threshold preset based on the objective attributes of the underlying physical units Compare: like Then determine the time. There is a risk of distorted prediction in the neural network, so these nodes are marked as power mutation nodes and assigned a directional compensation flag. =1; like If it is a stable node, it is marked as such and assigned a directional compensation flag. =0; At the power mutation node, a preset closed-loop feedback gain coefficient is introduced to convert the power deviation rate of change into a reverse compensation power term. ; In the formula, This represents the reverse compensation power term converted to the power dimension; This represents the preset closed-loop feedback gain coefficient (in the actual virtual power plant operation scenario, the preset closed-loop feedback gain coefficient). The value is highly correlated with the communication delay and control response dead zone of the underlying dominant equipment. To ensure the absolute stability of the compensation system, the closed-loop feedback gain coefficient... The typical value range is between 0.5 seconds and 3.0 seconds. This represents the rate of change of the average power deviation obtained from the preceding calculations; It should be noted that, in order to ensure the consistency of the physical dimensions on both sides of the equation, the closed-loop feedback gain coefficient... The dimension is time ( Its physical essence is equivalent to the differential lead time constant of the feedback control domain, making the converted... It has standard power dimensions.

[0034] The initial predicted power value corresponding to the power mutation node is compensated by reverse compensation power (the reverse compensation power term is physically superimposed on the initial predicted value marked as the mutation node to forcibly suppress the overshoot or follow-lag phenomenon of the neural network in the range of drastic power changes), and the compensated sequence is smoothly reconstructed to finally obtain the node value of the target output prediction sequence: ; in, ; In the formula, This indicates the output of the future number after dynamic tracking error elimination. The node values ​​of the target output prediction sequence at each time step; Indicates the trigger time of the current prediction task; Indicates time The assigned directional compensation flag (set to 1 when the moment is determined to be a power mutation node, and set to 0 when it is determined to be a stable node). This represents a reverse compensation power term that incorporates the divergence trend of the closed-loop error from the previous scheduling cycle and transforms it into a standard power dimension. Through the above flags Gated isolation performs physical corrections to suppress network overshoot or follow hysteresis only in areas of drastic power changes; Finally, the discrete point set Perform smooth spline fitting and reconstruction to output the final target output prediction sequence that eliminates dynamic following error.

[0035] The virtual power plant generation capacity prediction and optimization system based on time-series feature enhancement also includes an optimization scheduling unit 4. The optimization scheduling unit 4 calculates the available adjustment margin of the virtual power plant in the future scheduling period based on the target output prediction sequence and the equipment operation boundary conditions of the target controllable load. Based on the equipment operating boundary conditions and available adjustment margins, and with the goal of minimizing scheduling operation costs, a multi-objective optimization model is constructed to solve and generate optimized scheduling instructions for each controlled device in the corresponding virtual power plant. The optimized scheduling unit 4 includes a margin calculation module 41 and a multi-objective optimization module 42; The margin calculation module 41 is used to receive the target output prediction sequence and simultaneously obtain the real-time equipment operation boundary conditions of each target controllable load inside the virtual power plant. The target output prediction sequence is used as the benchmark output evolution trajectory and is physically superimposed and optimized spatial intersection calculated with the equipment operation boundary conditions to calculate the available adjustment margin of the virtual power plant in the future scheduling period (including the upward adjustment margin and the downward adjustment margin). Among them, the equipment operation boundary conditions include at least the absolute power amplitude boundary, the dynamic ramp rate boundary, the energy state and process tolerance boundary, and the start-stop cycle and service life boundary. In this embodiment, the absolute power amplitude boundary is used to define the absolute physical range within which the active power of the device can be adjusted downwards or upwards when participating in virtual power plant dispatch. The basis and scope for its definition are as follows: For controllable loads such as heavy-duty fans or pumps driven by variable frequency drives (VFDs), excessively low speeds can lead to fluid stall or insufficient cooling, while excessively high speeds can cause motor overload. Therefore, the lower limit is typically not zero, but rather set as the minimum guaranteed power to maintain basic process operation. In typical industrial scenarios, the power amplitude boundary is limited to between 40% and 100% of the equipment's rated power (i.e.,...). and The power value calculated by the scheduling command must strictly fall within this closed interval.

[0036] The dynamic ramp rate boundary is used to limit the maximum rate at which power can change dramatically between two adjacent scheduling time steps, in order to prevent mechanical shaft torque overload or thermal shock to electrical components. Its definition and scope are as follows: Based on the physical nameplate parameters of the equipment, for large mechanical rotating loads, the maximum allowable physical ramp rate threshold is typically defined in the range of 0.1MW / s to 1.5MW / s; for some electric heating loads without inertia equipped with solid-state relays, the ramp rate boundary can be relaxed to a full capacity step per second.

[0037] Energy state and process tolerance boundaries are used to define the state of charge (SOC) or thermodynamic state boundaries of the physical medium for loads with energy storage / buffering characteristics (such as electrochemical energy storage, central air conditioning thermal storage, and centralized electric boiler thermal storage). The basis and scope for their delineation are as follows: For electrochemical energy storage devices, to prevent the battery from undergoing accelerated lifespan degradation due to deep charging and discharging, the SOC operating boundary is strictly defined within the range of 20% to 80%. For HVAC or temperature-controlled loads, the process tolerance boundary is determined by the safe dead zone of the workshop ambient temperature, typically defined as ±1.5℃ to ±2.5℃ from the set process target temperature. When the state parameters touch this boundary, the system will forcibly deprive the device of its regulatory response qualification.

[0038] The start-stop cycle and operational lifespan boundaries are used to prevent virtual power plant cloud platforms from frequently issuing start-stop commands to underlying equipment in pursuit of short-term economic optimization, which could lead to the burnout of large contactors or mechanical wear. The basis and scope for defining these boundaries are as follows: For megawatt-level dominant loads, once they are scheduled to start or shut down, the minimum continuous operating time boundary is usually defined as 15 to 60 minutes; at the same time, the maximum number of times a single device is allowed to participate in large-scale cross-condition scheduling within a natural day is defined as 5 to 10 times.

[0039] It is worth noting that, in this embodiment, the specific steps involved in calculating the available adjustment margin of the virtual power plant during the future dispatch period are as follows: Suppose that the virtual power plant contains a total of A target controllable load, using an index express( ); for the future One prediction time step (i.e., time) First, extract a single device. Physical boundary conditions; by introducing dynamic availability flags. To integrate discrete state constraints, and to integrate continuous amplitude and ramp constraints through the extreme value intersection function, the calculation of a single device is performed. The actual bound at that moment and actual lower bound : ; ; In the formula, Indicates target controllable load In the future The upper limit of the actual adjustable power after being constrained by multiple physical boundaries; Indicates target controllable load In the future The lower bound of the actual adjustable power after being constrained by multiple physical boundaries; This indicates the upper limit of the absolute power amplitude boundary of the device; This indicates the lower limit of the absolute power amplitude boundary of the device; This indicates the reference state power of the device at the previous time step; This indicates the boundary of the upward dynamic ramp rate of the equipment; This indicates the downward dynamic ramp rate boundary of the equipment; Indicates the time step; and Represents the minimum and maximum functions in the intersection calculation of the optimization space; Indicates device The dynamic availability flag, in this embodiment, when the device... At any moment When the state simultaneously satisfies the energy state and process tolerance boundaries (such as not breaking through the SOC limit or temperature dead zone) and the start-stop cycle and operating life boundaries (such as satisfying the minimum continuous operating time), The value is 1; if any red line of a status is touched, then A forced setting to 0 indicates that the device loses its regulatory qualification at that moment.

[0040] It is worth noting that in this embodiment, the dynamic availability flag is introduced to reduce the complex and nonlinear discrete physical state constraints of the underlying equipment (such as the energy storage limit of the medium, the start-stop fatigue threshold, etc.) into binary mathematical multipliers. This not only gives the underlying target controllable load physical safety a veto right in terms of control mechanism, effectively preventing the risk of equipment damage caused by the top-level optimization algorithm excessively pursuing economic benefits, but also greatly eliminates the nonlinear constraint dimension when constructing the subsequent optimization scheduling model at the mathematical level, significantly improving the computational efficiency and model stability of large-scale virtual power plant clusters when solving scheduling instructions.

[0041] Furthermore, after completing the intersection calculation of all individual devices, the adjustable boundary of the target controllable load of the entire network is linearly aggregated to calculate the future time of the entire virtual power plant cluster. Overall physical operation upper bound Lower bound of overall physical operation : ; ; In the formula, This indicates the entire virtual power plant cluster at a future moment. The overall physical operation upper bound; This indicates the entire virtual power plant cluster at a future moment. The overall physical lower bound; This represents the total number of units within the virtual power plant that represent the target controllable load. The traversal index variable represents the target controllable load; This indicates all within the virtual power plant Perform traversal, summation, and aggregation operations on each target controllable load; In this embodiment, the virtual power plant is finally calculated at future time points. Real-world usable power adjustment margin With reduced power regulation margin : ; ; In the formula, This indicates the maximum upward adjustment margin of the system's power regulation, which can be increased beyond the baseline output and supported by physical boundaries. This indicates the maximum downward adjustment margin of the power regulation that the system can still reduce the voltage drop below the baseline output. The function is used to hedge extreme forecast biases, ensuring that the physical adjustment margin is always non-negative in mathematical expression.

[0042] The multi-objective optimization module 42 is used to receive the available adjustment margin and use it as the underlying hard safety constraint; it constructs a multi-objective optimization model with the objective function of minimizing the overall scheduling and operation cost of the virtual power plant (such as comprehensively considering the cost of electricity purchase, equipment depreciation cost and peak-valley arbitrage income); it uses a mathematical programming algorithm to solve the multi-objective optimization model in a rolling manner, calculates the optimal power allocation strategy, and converts the strategy into an optimized scheduling instruction for each controlled device in the virtual power plant for issuance.

[0043] In this embodiment, the optimal power allocation strategy is calculated, and the specific steps involved are as follows: Based on the preset total prediction window length, a multi-dimensional control time axis is established. The power dispatch command values ​​issued by each target controllable load in the virtual power plant at each time node within the future prediction window are used as the global decision variables of the optimization system. The objective function is to minimize the overall net scheduling and operation cost of the virtual power plant within the entire prediction window. Define the total step size of the future prediction window as The current prediction task is triggered at the following time. The future prediction step size number is ( ); Let the index of the target controllable load be ( ,in (Total number of devices), the global decision variable is defined for a single device. In the future The issued power scheduling command value .

[0044] Objective function for constructing a multi-objective optimization model The specific calculation formula is as follows: ; In the formula, This represents the overall net cost of scheduling and operation of the multi-objective optimization model within the entire future prediction window (objective function value). This indicates the preset total prediction window length (i.e., the total number of prediction steps). Indicating virtual power plants in the future The total basic electricity purchase cost across the entire network; This indicates that the various devices within the virtual power plant will be available at future times. The resulting depreciation costs; Indicating virtual power plants in the future Peak-valley arbitrage and ancillary service revenue obtained through peak shifting and valley filling; It should be further noted that in actual industrial virtual power plant day-ahead or intraday rolling dispatch scenarios, in order to fully cover a complete thermodynamic charge-discharge cycle of energy storage or temperature-controlled loads, and to avoid computational disasters caused by excessively long windows, if the time step is... The value is 15 minutes, and the total prediction window length is... The typical value range is usually set between 16 and 96 (that is, the forward scheduling window for the next 4 to 24 hours).

[0045] Among them, basic electricity purchase cost Peak-valley arbitrage profits The coupling calculation term is: ; Equipment operating depreciation cost The calculation items are: ; In the formula, This indicates the time step for discrete data sampling and control distribution; Indicates time The time-of-use electricity price or real-time spot market electricity price published by the power grid; This represents the total number of units within the virtual power plant that represent the target controllable load. The traversal index variable represents the target controllable load, and ; This represents the global decision variable representing the decision output of a multi-objective optimization model, i.e., the controllable load for a single target unit. In the future The issued power dispatch command value; by multiplying the time-of-use price and the real-time dispatch power, the electricity consumption penalty during high-price periods and the energy storage profit during low-price periods are automatically coupled into the same scalar model; Indicates the target controllable load of a single unit At the previous time step (i.e., time point) The power scheduling command value; This represents the absolute power change of the equipment in adjacent time steps. This term is used to forcibly suppress the extremely frequent bidirectional adjustments of the underlying equipment caused by the solver in pursuit of extreme electricity price differences, and to mathematically quantify the implicit depreciation cost of the equipment. Specifically, Indicates based on target controllable load The equipment wear characteristics are pre-set with a unit power fluctuation depreciation penalty factor, the range of which is: For high-loss-sensitive devices (such as electrochemical energy storage systems), cell lifespan is extremely sensitive to the depth and frequency of charge and discharge; each power reversal is accompanied by substantial chemical degradation. Therefore... The typical value range is the highest, usually calibrated between 100 yuan / MW and 500 yuan / MW; For moderately loss-sensitive equipment (such as heavy-duty fans, water pumps, and industrial compressors driven by frequency converters), power fluctuations primarily cause torque fatigue in the mechanical shaft system and thermal stress losses in the frequency converter's IGBT modules; therefore... The typical value range is moderate, usually between 10 yuan / MW and 80 yuan / MW. For low / zero-loss sensitive equipment (such as centralized electric boilers and electric heating furnaces controlled by solid-state relays without contact), power regulation is purely based on resistance heating, with almost no mechanical wear and extremely low lifespan degradation; therefore, The typical range of values ​​is the lowest, usually calibrated between 0 yuan / MW and 5 yuan / MW (or even its fluctuating depreciation is ignored directly in the model).

[0046] Furthermore, to ensure that the scheduling instructions generated by the optimization algorithm in the mathematical space are absolutely safe and executable when mapped to the underlying physical devices, the system nests two layers of hard constraints around the above objective function: Underlying physical constraints: The boundary constraints are combined with the actual bounds of the individual device output by the margin calculation submodule 41. and actual lower bound As the first layer of hard inequality constraints in the optimization space, instructions are prohibited from exceeding physical limits: ; in, ; Top-level scheduling constraints: Extract the total target output scheduling demand instructions issued by the upper-level power grid to this virtual power plant. This is used as an equality constraint for cluster power balancing, forcing the sum of cluster power commands to match scheduling requirements: ; In the formula, This indicates the time period issued by the higher-level power grid to the virtual power plant. The total target output scheduling requirement instruction (i.e., the target value of the cluster power balance equation constraint); After completing the construction of the objective function and the binding of constraints, since there is a nonlinear penalty term in the objective function that represents the absolute value of equipment fluctuation depreciation, the system first transforms the nonlinear model into a standard mixed integer linear programming (MILP) or quadratic programming (QP) paradigm by introducing internal auxiliary variables. Subsequently, the system invokes the built-in mathematical programming algorithm to solve the problem at the current moment. The optimal power allocation sequence matrix covering the entire future prediction window is calculated in a single forward pass. .

[0047] To effectively resist random disturbances in the external environment and the cumulative drift error of the AI ​​prediction model in the future, a rolling time window error-proofing mechanism of model predictive control (MPC) is adopted: Only the scheduling instruction corresponding to the first future time step (i.e., time step) in the optimal power allocation sequence matrix is ​​extracted. of The system will then use this as the final decision and send it to the controllable load control terminals of each target for execution. When the system time advances to the next real time step, the system will re-collect the latest field status and slide the entire prediction window forward, repeating the closed-loop optimization process from target construction to the first step of issuance.

[0048] The virtual power plant generation capacity prediction and optimization system based on time-series feature enhancement also includes a closed-loop correction unit 5. The closed-loop correction unit 5 is used to issue and execute optimization scheduling instructions, monitor the actual output response data of the virtual power plant in real time, and feed the actual output response data back to the prediction generation and error compensation module to update the power deviation change rate on a rolling basis.

[0049] In this embodiment, the closed-loop control logic and execution steps of the closed-loop correction unit 5 are as follows: Extract the optimization scheduling instruction for the first future time step output by the multi-objective optimization module 42 based on the rolling time window mechanism. The instruction is represented as an abstract mathematical matrix within the system. The closed-loop correction unit 5 parses and encapsulates the mathematical instruction into a low-level control message that conforms to the communication standards of the underlying controlled devices (such as Modbus-TCP, IEC 61850, or MQTT protocol) through the built-in industrial IoT communication gateway. Subsequently, the message is sent point-to-point to the local controllers (PLC or DCS systems) of each target controllable load in the virtual power plant, driving the circuit breakers, frequency converters, or contactors of the physical devices to perform corresponding power regulation actions. After control commands are issued, due to inherent communication delays, mechanical inertia ramp-up, and control dead zones in physical equipment, the actual power response trajectory of the equipment often cannot perfectly match the issued dispatch commands. Therefore, by deploying smart meters and high-speed data acquisition terminals at the grid connection points of each controlled device, the actual power output data of each device in the current dispatch cycle is monitored and collected in real time at extremely high sampling frequencies of seconds or milliseconds, and then aggregated upwards to form the overall actual output response data sequence of the virtual power plant. The closed-loop correction unit 5 then transmits the latest actual output response data sequence to the front-end dynamic prediction and compensation unit 3 via a high-speed data bus. Upon receiving the real observation data, the dynamic prediction and compensation unit 3 uses it as a historical known condition for the latest scheduling cycle and re-substitutes it into the dynamic error sequence solution formula. ,in, This indicates that the system, within the previous scheduling cycle, corresponds to the [number]th [time / period]. The dynamic power deviation value generated by each discrete sampling point, the positive or negative value of which represents the drift direction of the model prediction (for example, when it is greater than 0, it means that the model prediction lags behind the actual power evolution, and when it is less than 0, it means that the model prediction is overshooting). This indicates that the virtual power plant was at a historical moment. Actual power observations Represents a historical time step index, and satisfies This value is obtained in real time by the closed-loop correction unit 5 through the underlying physical sensors, and is used as input as known objective physical facts. This indicates that the system, in the previous scheduling cycle, targeted historical moments. Historical predicted power values ​​are pre-output and recorded, serving as a benchmark for identifying systematic errors in the model's dynamic following process through subtraction. The dynamic error sequence is obtained by calculating the actual physical execution deviation of the system in the just-past scheduling cycle using the dynamic error sequence solution formula. Based on the dynamic error sequence obtained from the solution, the average power deviation change rate is then calculated and updated. .

[0050] The updated It will be directly used as a feedforward physical correction term to correct the initial output prediction sequence in the next prediction window, thereby realizing feedforward physical correction and global closed-loop self-healing of the initial output prediction sequence in the next prediction window.

[0051] In summary, the present invention proposes a virtual power plant power generation capacity prediction and optimization system based on time-series feature enhancement. This addresses the issue that existing optimization software, in pursuit of maximizing the electricity price difference on paper, blindly and frequently demands deep charging and discharging of energy storage cells, or issues dozens of frequent start-stop commands to heavy-duty wind turbines within a single day. As a result, the meager electricity price difference earned by the virtual power plant cannot compensate for the enormous equipment replacement costs caused by contactor burnout, severe bearing wear, and accelerated battery degradation. Therefore, the present invention, in the optimization scheduling unit 4, introduces the "depreciation penalty coefficient" and "absolute power change amplitude" into the economic optimization objective function. This empowers the solver to automatically calculate costs and prioritize the use of lossless electric boilers to respond to commands. Only when the grid price difference is large enough to cover the lifespan depreciation cost of the energy storage batteries will a dispatch order be issued. This achieves the ultimate balance between the electricity revenue of the virtual power plant and the protection of heavy assets.

[0052] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A virtual power plant generation capacity prediction and optimization system based on time-series feature enhancement, characterized in that, include: Multi-source data acquisition and filtering unit (1) is used to acquire the network-wide historical power sequence of the virtual power plant in real time, and filter out the target controllable load from the virtual power plant according to the preset capacity ratio threshold, and collect the real-time operating status parameters of the target controllable load; The time-series feature enhancement unit (2) calculates the system equivalent time delay parameters of the target controllable load based on the real-time operating status parameters, filters and reconstructs the historical power sequence of the entire network, and generates an enhanced time-series feature sequence. The dynamic prediction and compensation unit (3) is used to input the enhanced time series feature sequence into the pre-trained time series prediction network and output the initial output prediction sequence; and by extracting the actual execution power and predicted power of the virtual power plant in the previous scheduling cycle, calculating the power deviation change rate between the two, and using the power deviation change rate to numerically compensate the initial output prediction sequence to obtain the target output prediction sequence. The optimization scheduling unit (4) calculates the available adjustment margin of the virtual power plant in the future scheduling period based on the target output prediction sequence, constructs a multi-objective optimization model with the equipment operation boundary conditions and available adjustment margin as constraints and the goal of minimizing scheduling operation costs, and solves to generate optimization scheduling instructions. The closed-loop correction unit (5) is used to issue and execute the optimized scheduling instruction, monitor the actual output response data of the virtual power plant in real time, and update the power deviation change rate based on the actual output response data.

2. The virtual power plant generation capacity prediction and optimization system based on time-series feature enhancement according to claim 1, characterized in that, The target controllable load is classified into heavy machinery rotation load and temperature control and heat processing load according to its power response time delay mechanism; Wherein, when the target controllable load is the heavy machinery rotation load, the real-time operating status parameters include at least the actual speed of the motor rotor, the stator side operating current, and the mechanical shaft end load torque; When the target controllable load is the temperature control and heat processing type load, the real-time operating status parameters include at least the current input power, the temperature difference between the supply and return water circuits of the circulating medium, and the deviation between the actual temperature of the target process environment and the set temperature.

3. The virtual power plant generation capacity prediction and optimization system based on time-series feature enhancement according to claim 2, characterized in that, The time-series feature enhancement unit (2) includes a parameter calculation and aggregation module (21) and a filtering and reconstruction module (22). The parameter calculation and aggregation module (21) calculates the load response time constant of the target controllable load based on the real-time operating status parameters, and calculates the equivalent time delay parameters of the system by weighting according to the load response time constant and the current actual power ratio of the target controllable load. The filtering and reconstruction module (22) is used to convert the system equivalent time delay parameters into first-order inertial filter coefficients, and to use the discrete differential filtering algorithm to iteratively smooth each time-series sampling node of the historical power sequence of the entire network, and output an enhanced time-series feature sequence.

4. The virtual power plant generation capacity prediction and optimization system based on time-series feature enhancement according to claim 3, characterized in that, The specific steps involved in calculating the load response time constant of the target controllable load are as follows: When the target controllable load is the heavy machinery rotation load, the load response time constant is calculated based on the dynamic ratio of the actual rotational speed of the motor rotor to the load torque at the mechanical shaft end, combined with the preset equipment rotational inertia equivalent; wherein, the load response time constant is positively correlated with the actual rotational inertia of the motor rotor and negatively correlated with the magnitude of the resultant force of the load torque at the mechanical shaft end and the driving electromagnetic torque. When the target controllable load is a temperature control and heat processing load, the corresponding load response time constant is determined based on the deviation between the actual temperature and the set temperature of the target process environment, the equivalent temperature difference ladder formed by the temperature difference of the supply and return water circuits of the circulating medium, and the preset equivalent heat capacity of the system and the current input active power of the core working element.

5. The virtual power plant generation capacity prediction and optimization system based on time-series feature enhancement according to claim 3, characterized in that, The specific calculation process for the enhanced temporal feature sequence is as follows: Collect the current actual power consumption of each target controllable load and calculate its actual power proportion in the original total power of the virtual power plant network. The actual power ratio is used as a dynamic weighting coefficient to perform a weighted summation of the calculated load response time constants. By integrating the pre-calibrated reference time constants of the remaining background devices in the entire network, the equivalent time delay parameters of the system are calculated; Using the system's equivalent time delay parameters as the smoothing constraint benchmark for the first-order inertial filtering algorithm, dynamic filtering coefficients are constructed by combining the system data sampling period. By using dynamic filtering coefficients to perform sliding iterative reconstruction on each discrete sampling point of the historical power sequence of the entire network, an enhanced time series feature sequence is finally obtained.

6. The virtual power plant generation capacity prediction and optimization system based on time-series feature enhancement according to claim 1, characterized in that, The dynamic prediction and compensation unit (3) includes an initial prediction module (31) and a dynamic compensation module (32). The initial prediction module (31) is used to input the enhanced temporal feature sequence into a pre-trained long short-term memory temporal prediction network, and to perform sequence mapping and feature decoding of the forward time step using a gating mechanism, and output the initial output prediction sequence. The dynamic compensation module (32) is used to extract the actual power sequence and the historical predicted power sequence of the virtual power plant in the previous scheduling cycle, and to calculate the power deviation change rate between the two using the discrete difference algorithm; Simultaneously, a sliding window threshold detection algorithm is used to identify power mutation nodes in the initial power output prediction sequence, and the power deviation change rate is used as a feedforward physical correction term to perform reverse cancellation and numerical compensation on the predicted values ​​at the power mutation nodes in the initial power output prediction sequence, and finally output the target power output prediction sequence.

7. The virtual power plant generation capacity prediction and optimization system based on time-series feature enhancement according to claim 6, characterized in that, In the dynamic compensation module (32), the specific calculation logic for obtaining the target output prediction sequence through numerical compensation is as follows: The difference between the actual execution power sequence and the historical predicted power sequence is calculated at discrete time steps to obtain the dynamic error sequence; The power deviation change rate is calculated by performing a discrete first-order difference operation on the dynamic error sequence. The initial output prediction sequence is traversed using a sliding observation window to calculate the slope of the power difference between adjacent prediction time steps; When the absolute value of the power difference slope is greater than the preset physical ramp rate threshold, the corresponding prediction time step is marked as a power mutation node. At the power mutation node, a preset closed-loop feedback gain coefficient is introduced to convert the power deviation rate of change into a reverse compensation power term. The initial predicted power value corresponding to the power mutation node is compensated by reverse power compensation, and the compensated sequence is smoothly reconstructed to finally obtain the target output prediction sequence.

8. The virtual power plant generation capacity prediction and optimization system based on time-series feature enhancement according to claim 1, characterized in that, The optimization scheduling unit (4) includes a margin calculation module (41) and a multi-objective optimization module (42). The margin calculation module (41) is used to receive the target output prediction sequence and simultaneously obtain the real-time equipment operation boundary conditions of each target controllable load inside the virtual power plant. The target output prediction sequence is used as the benchmark output evolution trajectory and is physically superimposed and optimized spatial intersection calculated with the equipment operation boundary conditions to calculate the available adjustment margin of the virtual power plant in the future scheduling period. The available adjustment margin includes the upward adjustment power adjustment margin and the downward adjustment power adjustment margin. The multi-objective optimization module (42) is used to receive the available adjustment margin and use it as the underlying hard safety constraint. It constructs a multi-objective optimization model with the objective function of minimizing the overall scheduling and operation cost of the virtual power plant. It uses a mathematical programming algorithm to solve the multi-objective optimization model in a rolling manner, calculates the optimal power allocation strategy, and converts the strategy into an optimized scheduling instruction for each controlled device in the virtual power plant for distribution.

9. The virtual power plant generation capacity prediction and optimization system based on time-series feature enhancement according to claim 8, characterized in that, The equipment operation boundary conditions include at least the absolute power amplitude boundary, the dynamic ramp rate boundary, the energy state and process tolerance boundary, and the start-stop cycle and service life boundary.

10. The virtual power plant generation capacity prediction and optimization system based on time-series feature enhancement according to claim 1, characterized in that, The closed-loop correction unit (5) updates the power deviation rate of change, and the steps involved are as follows: Extract the optimization scheduling instruction of the first future time step output by the multi-objective optimization module (42), parse its protocol and convert it into a control message that conforms to the communication standard of the underlying controlled device, and drive the corresponding physical device to perform actual power adjustment actions; By deploying intelligent acquisition terminals at the bottom layer, the actual power data of the virtual power plant in the current scheduling cycle is collected in real time at a preset sampling frequency, and aggregated to form an actual output response data sequence; The actual output response data sequence is transmitted back to the dynamic prediction and compensation unit (3); the dynamic prediction and compensation unit (3) extracts the actual execution power observation value in the previous scheduling cycle, uses it as the historical known condition of the latest scheduling cycle, and calculates the actual execution power observation value and the corresponding historical predicted power value to obtain the dynamic error sequence used to characterize the physical execution deviation; Based on the dynamic error sequence, the dynamic prediction and compensation unit (3) is driven to continuously update the average power deviation change rate, thereby realizing the feedforward physical correction and global closed-loop self-healing of the initial output prediction sequence in the next prediction window.

Images