Intelligent scheduling method for comprehensive virtual power plant

Abstract

The invention discloses an intelligent scheduling method for a comprehensive virtual power plant, and relates to the technical field of power plant intelligent scheduling, and the method comprises the steps: carrying out the redundancy elimination filtering and unified time sequence calibration of multi-node sensing data of a cold and hot energy station, energy storage equipment and the like through the deployment of an edge side high-concurrency collection module, generating a high-quality time sequence fusion sequence, and screening schedulable resources; on-line model correction is realized through an error dynamic evaluation mechanism, a multi-stage scheduling strategy is generated by adopting an improved particle swarm algorithm aiming at the targets of economy, reliability and environmental protection, and an optimal instruction is output in combination with an equipment constraint condition. Through priority control and real-time monitoring index evaluation deviation, a rolling correction and standby strategy is triggered, equipment-level fault risk quantification and self-healing control are realized, abnormity is rapidly isolated, and redundant resources are started. Through data-prediction-scheduling-execution closed-loop cooperation, the resource utilization rate, the scheduling precision and the system fault-tolerant capability are significantly improved, and the scheduling capability of source heterogeneous resources is enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent dispatching in power plants, and specifically to an intelligent dispatching method for an integrated virtual power plant. Background Art

[0002] Under the background of today's energy transformation and the deepening reform of the power system, with the rapid development of distributed renewable energy (such as photovoltaic and wind power) and the continuous maturity of energy storage and adjustable load technologies, the traditional power grid form dominated by centralized power sources is gradually evolving towards a multi-source collaborative integrated energy network. In application scenarios such as industrial parks, large commercial complexes, and urban energy microgrids, multiple energy forms (electricity, cooling, heating, etc.) and numerous decentralized devices (cold and heat energy stations, distributed photovoltaics, energy storage devices, adjustable loads, etc.) coexist, each with different operating characteristics and dispatchable characteristics. The concept of a virtual power plant (VPP) is based on Internet of Things, intelligent prediction, and dispatch optimization technologies to converge data and conduct collaborative management and control of the above heterogeneous resources, thereby realizing in-depth excavation and unified dispatch of distributed energy. While meeting the energy demand and system security, it further improves the utilization rate of clean energy and economic benefits. Due to the dual drive of the continuous improvement of the external power market mechanism and the diversification of internal electricity demand, virtual power plants have become increasingly prominent in the new power system and have become an important starting point for building a high-proportion clean energy power grid.

[0003] In a Chinese invention patent with the application publication number CN117439171A, an intelligent dispatching method, system, and medium based on a virtual power plant are disclosed, which solve the deficiencies of the prior art. The method includes steps of obtaining historical electricity consumption data on the user side, predicting the future electricity demand on the user side based on the historical electricity consumption data on the user side; the virtual power plant formulating a power generation plan according to the future electricity demand on the user side, determining the proportion of power generation of multiple power generation energy sources according to the future power generation demand; the virtual power plant supplying power to the user side according to the power generation plan, and at the same time, the virtual power plant monitoring the total power generation of multiple power generation energy sources and the electricity demand on the user side. If the difference between the total power generation of multiple power generation energy sources and the electricity demand on the user side exceeds a threshold, the proportion of power generation of multiple power generation energy sources is adjusted in real time to ensure the actual usage demand of the total power generation of multiple power generation energy sources.

[0004] However, combining the above prior art and actual application scenarios: In the large-scale coupling of multi-source heterogeneous devices and dynamic operating environments, the key technical problems to be solved urgently are: How to balance real-time data fusion, high-precision prediction, and generation of feasible and reliable scheduling strategies for various distributed resources within a multi-type energy network with high concurrency and strong uncertainty, and be able to perform fast self-healing control in case of equipment failures or sudden fluctuations. Specifically, when the cold and heat energy stations, adjustable loads, energy storage, and photovoltaic systems generate power fluctuations or anomalies during operation due to meteorological changes, electricity price fluctuations, or equipment aging, without a unified data fusion mechanism and dynamic resource identification means, the system will be difficult to timely grasp the true operating conditions of the equipment, resulting in a decline in the accuracy of the prediction model and making the multi-objective scheduling strategy unable to be effectively implemented. At the same time, in case of failures, such as energy storage failures, grid transients, or large fluctuations in distributed power output, if immediate diagnosis cannot be carried out and quickly switched to an alternative scheduling plan, it is easy to cause local supply gaps or energy waste, and even induce greater energy supply security risks.

[0005] To this end, the present invention provides an intelligent scheduling method for an integrated virtual power plant. Summary of the Invention

[0006] (I) Technical Problems to be Solved Aiming at the deficiencies of the prior art, the present invention provides an intelligent scheduling method for an integrated virtual power plant. By deploying a high-concurrency acquisition module on the edge side, redundant filtering and unified time-series calibration are performed on multi-node sensing data such as cold and heat energy stations and energy storage devices to generate a high-quality time-series fusion sequence and screen schedulable resources. Through an error dynamic evaluation mechanism, online correction of the model is realized. For economic, reliability, and environmental protection objectives, an improved particle swarm optimization algorithm is used to generate multi-stage scheduling strategies, and the optimal instructions are output in combination with equipment constraint conditions. By evaluating the deviation of the priority control and real-time monitoring indicators, rolling correction and backup strategies are triggered to realize the quantification of equipment-level failure risks and self-healing control, quickly isolate anomalies, and start redundant resources. Through the closed-loop coordination of data-prediction-scheduling-execution, the resource utilization rate, scheduling accuracy, and system fault tolerance are significantly improved, thus solving the technical problems described in the background art.

[0007] Under the overall framework of the virtual power plant, a systematic solution covering data acquisition and fusion, prediction modeling, scheduling optimization and closed-loop coordination, and fault diagnosis and self-healing control is constructed to ensure that various energy resources can still be coordinated and managed under complex working conditions, taking into account economic, reliability, and environmental protection objectives, and improving the ability to respond promptly to sudden failures or abnormal operations.

[0008] (II) Technical Solutions To achieve the above objectives, the present invention is realized through the following technical solutions: An intelligent scheduling method for an integrated virtual power plant, including, When multi-source sensing nodes report raw data When it is time, the sensor data of distributed power equipment is aggregated by deploying a high-concurrency acquisition module on the edge side, and preprocessed data is generated after preprocessing with a message queue and a timestamp. , unified time series calibration is performed, and cleaned data is generated. , through dynamic weights and attenuation factors , an exponentially weighted fusion algorithm generates a time series fusion sequence. , combined with the equipment operating conditions and environmental data to construct a comprehensive feature vector. , a discriminant function and a decision threshold are used to generate the resource status. ; When the time series fusion sequence and the resource status are obtained, the initial feature vector is decomposed by wavelet transform at multiple scales to generate a prediction input vector. , after parallel training, load prediction , power generation prediction and energy storage prediction are generated. The error function is used to dynamically evaluate the prediction deviation, and online parameter correction is triggered when the error exceeds the error threshold; When the prediction result is input with the resource status , the multi-objective optimization objective function integrates the constraint objectives, and the improved particle swarm algorithm is used to optimize the scheduling decision vector X in multiple stages. A hard power balance constraint is constructed based on the rated parameters of the equipment, and an optimal scheduling scheme is generated using the adaptive weight adjustment mechanism. , and a device execution plan is generated. The effectiveness of the scheme is evaluated by decomposing the economic-reliable-environmental protection indicators, realizing multi-objective dynamic trade-off and coupling of global-local scheduling strategies; When the optimal scheduling scheme is sent to the on-site equipment, the instruction is converted into a device-level operation sequence by the on-site energy management system, and the control instruction is sent synchronously using the priority allocation mechanism and the monitoring index vector is collected. A performance metric function is constructed to evaluate the execution deviation. When the deviation accumulation exceeds the performance threshold , a rolling correction process is triggered, and a local optimal solution is generated by re-optimizing using a local time window. ; When the monitoring index vector and the time series fusion sequence are abnormal, a fault determination function , identify device-level fault modes. When the fault determination function exceeds the fault threshold , trigger self-healing control. Adopt a backup scheduling plan to start redundant devices and adjust the load distribution strategy. After repair, based on the safety threshold restore the device scheduling authority.

[0009] Preferably, deploy sensors at distributed power equipment. The original data reported by each node at time is recorded as . The original data and the corresponding timestamp will be initially aggregated on the edge side; Filter and remove redundancy from the original data to obtain preprocessed data after preliminary screening ; perform synchronous acquisition and unified time series alignment for different nodes; Preferably, import the preprocessed data into the central database. After preprocessing, obtain the cleaned data . Perform weighted fusion on the data of multiple nodes to generate a unified single-point fusion value at a moment : Arrange the single-point fusion values at different times in sequence to form a time series fusion sequence ; ; Preferably, integrate the time series fusion sequence with the device operating condition records and meteorological data to form a comprehensive feature vector : Analyze the comprehensive feature vector to determine whether the key resources are in a schedulable or abnormal state, and define a determination function :

[0010] where is the determination threshold, used to comprehensively evaluate factors such as resource health and adjustability; Unify the schedulable state and operation mode of each distributed power equipment at time and record them as the resource state Preferably, select key data from the time series fusion sequence and the resource state and combine with external inputs to construct an initial feature vector . The initial feature vector Perform wavelet transform on some of the time series components in it to extract high-frequency perturbations and low-frequency trends, decompose the original time series into several multi-scale components; merge the multi-scale components obtained by wavelet transform with other key features to form a prediction input vector ; Preferably, after obtaining the prediction input vector Perform short-term prediction based on a deep neural network and ultra-short-term prediction using a lightweight time series model, and output key predicted quantities within different prediction time durations respectively; Train the neural network model in a rolling window manner and perform cross-validation, and obtain a scheduling prediction model after dynamically adjusting the learnable parameters; Preferably, after the scheduling prediction model is put into online operation, obtain the actual measurement values at high frequency , and compare with the model output ; Define a custom error function to dynamically measure the prediction accuracy as follows:

[0011] In the formula: is the target index, , is the number of prediction objects to be evaluated simultaneously; , where: represents the predicted value of the th target, is the corresponding actual observed value; is the relative time after the current evaluation start time , and the value range is ; is the derivative error weight corresponding to the th prediction target; represents the evaluation duration; is the time decay factor of the th target; According to the fluctuation of the error function , if the error continuously exceeds the preset error threshold, trigger the online correction program; Incrementally incorporate the latest prediction input vector and the target predicted value acquired in real time into the model training set, and update the model parameters through small-scale iterative training, and dynamically adjust the weight matrix , bias vector and dynamic weight coefficient etc. to complete online calibration; Preferably, according to the short-term / ultra-short-term prediction results of the key predicted quantities, combined with the resource status , construct a multi-objective optimization objective function ; With the hard constraints of the device rated operating range scheduling decision vector X, referring to the prediction range and time resolution, power balance and safety margin should be ensured in each discrete scheduling period; Preferably, the particle swarm position represents the scheduling decision vector X, and the fitness is defined as the multi-objective optimization objective function to achieve minimization by taking the opposite number or other means; taking the system state within the scheduling period as the state space and the scheduling decision vector X as the action space, and using the reward or punishment corresponding to the multi-objective optimization objective function to learn the policy; adjusting and setting the initial parameters of the algorithm according to the device topology information, prediction duration, and time granularity; The multi-objective optimization solution process determines convergence by monitoring the iterative changes of the multi-objective optimization objective function and the satisfaction degree of the feasible solution constraints. If full convergence cannot be achieved within the preset number of iterations or time, the current optimal feasible solution is taken, and finally the optimal scheduling plan is output ; Map the output optimal scheduling plan to the actual device level, format and output the scheduling policy, and decompose the performance of each item in the optimal scheduling plan in the objective function; Preferably, according to the optimal scheduling plan , extract the operating power, start-stop status, and charge-discharge plan of distributed power equipment in each period respectively, and convert them into specific execution instructions for the equipment; the on-site energy management system matches the execution points one by one according to the corresponding resource identifiers; Combined with the virtual power plant cloud platform, formulate a priority order for the operation instructions of each type of equipment; synchronously issue the instructions at the same moment or within the same scheduling cycle. After the on-site energy management system receives the instructions and completes the execution, it transmits the actual operation information of the equipment back to the cloud platform, which together with the data obtained by the sensors constitutes the comprehensive operation monitoring source; Preferably, define the monitoring index vector according to the prediction result, and introduce the performance metric function to measure the deviation degree between the execution effect of the optimal scheduling plan , and the formula is as follows:

[0012] In the formula: , represents the difference metric function between the current execution deviation or execution result and the scheduling plan at time ; is the executed scheduling instruction, is the real-time monitoring index vector; , a function for measuring scheduling deviation; In when, the deviation is amplified superlinearly so that large errors account for a higher proportion in the overall evaluation: when it returns to the linear case; In is the weight factor of the derivative term, and are both control parameters; In , represents the time decay factor; is the evaluation window length; If the performance metric function is above the preset performance threshold or a non - linear jump of the energy storage / load device is detected, a warning signal is sent to the cloud platform, and at the same time, the rolling correction process can be triggered; Preferably, when the performance metric function exceeds the performance threshold for several consecutive monitoring periods, the scheduling re - evaluation process is automatically triggered, integrating the latest data collected in real - time to form the state vector at the current moment; Use the improved particle swarm algorithm or deep reinforcement learning to quickly re - optimize the scheduling scheme within a local time window to obtain a new local optimal solution Compared with the original optimal scheduling scheme , if it can reduce the recent operation deviation or reduce costs / emissions, the original scheduling instruction can be locally replaced, and the original scheme is still used for the remaining periods. The locally optimal solution after rolling correction enters the instruction issuing process again and is subject to real - time monitoring and evaluation; Preferably, combining the time - series fusion sequence and the monitoring index vector is uniformly denoted as the comprehensive feedback data , and define the following fault determination function :

[0013] where: represents the measurement value of the fault risk at time ; is the possible fault mode or the number of leaf nodes, represents the fault mode index; is the weight coefficient related to the fault mode , is the recognition function of the time - series deep - learning model for the input data, represents for the fault mode Set of trainable parameters; Indicates at time For the fault mode Trigger duration or severity factor of Indicates the fault attenuation factor; When the fault determination function Exceeds the preset fault threshold It is determined that there is a high fault risk at present, the corresponding device is regarded as a faulty device, and a fault diagnosis instruction is issued; Preferably, when the fault determination function Exceeds the preset fault threshold When there is a rapid upward trend, immediately enter the self-healing control process; Re-evaluate the faulty device or associated load, and transmit an instruction to reduce or prohibit the scheduling of the faulty device, or increase the corresponding safety factor; suspend or terminate the issuance of the scheduling instruction for the faulty device, and call the following alternative scheduling scheme; After receiving the fault diagnosis instruction, immediately switch to this alternative strategy to quickly adapt to the changes in the load and energy supply structure after the fault. It can perform secondary optimization on non-faulty devices within a local time window to generate a temporary scheduling scheme; After the self-healing process is started, maintain or isolate the faulty device for a period of time, and gradually observe whether the risk value given by the fault tree model drops significantly: If , where: Is the safety threshold; that is, it indicates that the device fault has been eliminated or the risk is controllable, and then the normal scheduling scheme can be restored; If the device repair is completed, it can be added to the scheduling scope again: if the repair is not completed, it remains in the isolated state to avoid affecting the operation of the overall system.

[0014] (III) Beneficial effects The present invention provides an intelligent scheduling method for an integrated virtual power plant, which has the following beneficial effects: By deploying a high-concurrency acquisition and distributed computing mechanism on the edge side, the sensing data of various devices such as energy storage, cold and heat energy stations, adjustable load devices, and photovoltaic systems are uniformly processed; on the one hand, it reduces the data island phenomenon and the load of the central server, and on the other hand, through edge pre-aggregation and dynamic identification and mining of the changing rules of device working modes, it provides an accurate and real-time operation portrait for subsequent prediction and scheduling.

[0015] Using big data analysis and deep learning algorithms, short-term and ultra-short-term predictions are made for load demand and distributed power equipment. At the same time, correlation analysis is carried out in combination with external factors such as the state of charge and price of energy storage, and meteorology. Not only is the coupling relationship between the cold and heat energy stations, distributed power sources and energy storage considered in feature extraction, but the model parameters are dynamically updated through an online correction mechanism, significantly reducing the prediction deviation. Therefore, during peak and valley periods or when meteorological mutations occur, the actual operating conditions can still be better tracked.

[0016] Multiple requirements such as economy, reliability, and environmental protection are integrated into a unified objective function, and with the constraints of energy storage capacity constraints and equipment safety margin limits, improved particle swarm or deep reinforcement learning algorithms are comprehensively used to generate multi-dimensional scheduling solutions. This scheduling solution can finely set the allocation of energy storage and load at different time scales to ensure the maximization of benefits, reduction of scheduling risks, and further improvement of the utilization rate of clean energy under the fluctuations of electricity prices, loads, and photovoltaics.

[0017] By monitoring the real-time changes of key operating indicators (such as energy storage SOC, load tracking error, etc.) and evaluation functions, an adaptive dynamic scheduling system is constructed. Once execution deviation or supply-demand imbalance is detected, local re-optimization is triggered, and instructions are timely updated and sent to each resource end. This process effectively reduces the impact brought by prediction errors or external disturbances, ensuring that the virtual power plant is in an efficient and controllable state during daily operation.

[0018] By combining the fault tree with time-series deep learning to measure the fault risk, and triggering the standby scheduling mode or quickly isolating the faulty equipment when the risk exceeds the threshold. Therefore, the benefits and environmental objectives can be balanced in normal scenarios, and a smooth transition can be achieved through temporary solutions in case of faults or abnormal conditions, avoiding large-scale energy supply interruptions caused by single-point failures.

[0019] In summary, by enhancing information accuracy through multi-source data fusion and dynamic recognition, by enhancing the perception of load and distributed power source fluctuations through short-term and ultra-short-term predictions, by flexibly weighing between economic and environmental demands through multi-objective optimal scheduling, by maintaining real-time balance through closed-loop collaborative optimization, and by ensuring system safety redundancy and rapid recovery with the help of fault diagnosis and self-healing mechanisms, the practicality and efficiency of the virtual power plant are greatly improved as a whole. Brief Description of the Drawings

[0020] Figure 1 It is a schematic flow diagram of the intelligent scheduling method of the comprehensive virtual power plant of the present invention. Detailed Embodiment

[0021] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0022] Please refer to Figure 1 , the present invention provides an intelligent scheduling method for an integrated virtual power plant, including Step 1: When the multi-source sensing nodes report the original data , the sensor data of the distributed power equipment is aggregated through the edge-side high-concurrency acquisition module deployed. After preprocessing with the message queue and timestamp, the preprocessed data is generated , the unified time series calibration is performed, and the cleaned data is generated . Through the exponential weighted fusion algorithm of the dynamic weight and the attenuation factor , the time series fusion sequence is generated . Combining the equipment working conditions and the environmental data , the comprehensive feature vector is constructed . Using the discriminant function and the decision threshold , the resource status is generated ; The content included in the above Step 1 is as follows: Step 101: High-concurrency acquisition and edge preprocessing Sensors are deployed at the cooling and heating energy stations, adjustable load equipment, energy storage, and photovoltaic systems. The original data reported by each node at time is recorded as , and its corresponding timestamp is recorded as . Here, represents the sensing node number. The original data and the corresponding timestamp will be initially aggregated on the edge side; Filter and remove redundancy from the original data . When the same node reports multiple times in a very short time and the data has no obvious difference, it can be considered as duplicate data and discarded; the preprocessed data after preliminary screening is obtained ; Using the message queue or high-concurrency data bus to achieve synchronous acquisition and unified time series alignment of different nodes; the timestamp can be aligned to the unified reference time axis after edge processing; During use, through parallel data acquisition and edge preprocessing, the load on the central server is significantly reduced, and the network latency is shortened. Compared with the traditional centralized acquisition mode, an edge filtering and redundancy removal mechanism is introduced, enabling preliminary cleaning and format unification of massive sensing data before uploading.

[0023] Step 102: Implementation of data cleaning and fusion algorithm Import the preprocessed data into the central database, mark and process the obviously abnormal or missing data, and obtain the cleaned data after preprocessing , if a node is offline or the data jump amplitude is too large, record it separately for subsequent operation and maintenance handling; Use the following formula to perform weighted fusion on the data of multiple nodes to generate a unified single-point fusion value at a moment :

[0024] where: is the number of nodes participating in the current fusion; is the dynamic weight of node , usually adjusted within the range of , determined based on factors such as the historical accuracy of the node or the device status; is the cleaned data of this node; is the node decay factor, with a value between 0 and 1, used to dynamically correct the data contribution degree according to the node aging or interference degree; is the exponential function, used to smooth and amplify or attenuate the data influence of different nodes; Arrange the single-point fusion values at different times in sequence to form a time-series fusion sequence , denoted as ; ; During use, through a custom weighted exponential decay function, it is possible to dynamically reduce the weight of nodes with high noise or low credibility, further improve the accuracy of data fusion, and the unified output time-series fusion sequence provides high-quality multi-source data metrics for subsequent resource identification and predictive analysis.

[0025] Step 103: Dynamic resource identification and state modeling Integrate the time-series fusion sequence with the device operating conditions record (such as energy storage SOC, cold / hot load on / off status) and meteorological data to form a comprehensive feature vector :

[0026] is the single-point fusion value; Indicates the device status (such as energy storage charge level, load start / stop flag, etc.); Indicates environmental data (such as outdoor temperature, light intensity, etc.).

[0027] Using a set discriminant function or lightweight machine learning model For the comprehensive feature vector Conduct an analysis to determine whether the key resources are in a schedulable or abnormal state, and define a decision function :

[0028] Where Is the decision threshold, Used to comprehensively evaluate factors such as resource health and adjustability; Unify the schedulable status and operation mode of each distributed power device (such as energy storage devices, adjustable load devices, cold and heat source stations, and distributed power sources) at time Record as the resource status To form a resource configuration file; If , it means that the resource has the potential to respond to external scheduling. If It is regarded as currently non-adjustable or requires additional maintenance attention.

[0029] When in use, comprehensively utilize the time series fusion sequence , device status And environmental data Improve the ability to characterize multi-dimensional coupling characteristics, effectively identify high-value scheduling resources and filter out faulty or unstable resources. Compared with traditional methods that rely on a single threshold, this dynamic identification mechanism can adaptively adjust the decision threshold and feature weights under different working conditions. Deeply couple the multi-source fusion results with multi-dimensional operation information to construct a more flexible comprehensive feature vector , through a configurable discriminant function Achieve refined identification and status marking, with stronger adaptability.

[0030] Step 2. When obtaining the time series fusion sequence And resource status , perform multi-scale decomposition on the initial feature vector Through wavelet transform to generate a prediction input vector , generate load prediction , power generation prediction And energy storage prediction After parallel training, use the error function Dynamically evaluate the prediction deviation, and trigger online parameter correction when the error exceeds the error threshold; The content included in the said Step 2 is as follows: Step 201, Feature Data Assembly and Multi-Dimensional Preprocessing From the time-series fusion sequence and the resource status select the key data that can reflect the real-time operation characteristics of the system (such as the fusion power value of distributed photovoltaic, the current state of charge of energy storage, the start / stop signals of heating and cooling load equipment, etc.), and construct an initial feature vector by combining external inputs such as environmental / meteorological data and market electricity prices ; Among them, the time-series fusion sequence represents the comprehensive measurement information (such as the weighted result of the output of the photovoltaic array and the load-side current) after the fusion processing in step one at time ; is used to indicate the dispatchability and operation mode of each resource; To capture the fluctuation characteristics of the load and distributed power sources at different time scales, wavelet transform can be performed on some time-series components in the initial feature vector to extract high-frequency perturbations and low-frequency trends, and define the following wavelet analysis formula:

[0031] Where: represents the target component at continuous time (which can come from any time-series data in the initial feature vector ); is the complex conjugate function of the mother wavelet; and respectively represent the translation and scale parameters (both are positive values), which can be adjusted within a certain range according to the prediction period requirements; Through this wavelet analysis, the original time series can be decomposed into several multi-scale components, which is convenient for separately processing the slow-varying trend and short-term fluctuations in the subsequent scheduling prediction model; Merge the multi-scale components obtained by wavelet transform with other key features (such as environmental temperature and humidity, energy storage state of charge, etc.) to form the final prediction input vector , and perform unified normalization processing on the prediction input vector (such as mapping the interval to or ), and retain the same feature order and labels; Deeply combining the time-series fusion sequence with the resource status helps to improve the accuracy of the scheduling prediction model in depicting the real dynamics of the system.

[0032] By means of time-frequency domain analysis such as wavelet transform, the fluctuation characteristics of the load and distributed power sources at different time scales can be effectively separated, improving the prediction input vector Recognizability; Step 202, Construction of Short-term and Ultra-short-term Scheduling Prediction Models After obtaining the prediction input vector , short-term prediction is realized based on a deep neural network (DNN). Meanwhile, an ultra-short-term prediction is realized using a lightweight time series model (such as a network based on the gated recurrent unit GRU), and the key prediction quantities within different prediction time durations are output respectively (load), (distributed generation power), and (energy storage state of charge trend), etc.; Wherein represents the prediction step length. For example, for short-term prediction, discrete moments within the range of 1 to 24 hours can be taken, and for ultra-short-term prediction, 15 minutes or shorter intervals can be taken.

[0033] More specifically, obtain the time series fusion sequence and historical load data over several cycles , distributed power source data and external meteorological data , and merge these elements into the prediction input vector and send it into the scheduling prediction model to generate short-term or ultra-short-term load and power source prediction values , and other results; The following gives a general neural network-based prediction output formula (only for exemplary expression, not traditional mean or simple error statistics):

[0034] Wherein: represents the prediction value of the target variable (load, power output, etc.); represents the number of parallel sub-modules or different channels in the network (for example, the aggregation of multiple hidden layer units); is the dynamic weight coefficient output by the th sub-module (which can be adaptively adjusted according to the characteristics of the current moment, and the value range can be within ), and can be regarded as a learnable parameter within the network; and respectively represent the weight matrix and bias vector of the sub-module for the prediction input vector at time ; represents a non-linear activation function, which is used to enhance the fitting ability of the model to complex relationships, such as ReLU, tanh, etc.; Outputs of multiple key prediction quantities for short-term and ultra-short-term models , , , which can be obtained separately through a similar structure as above, and then weighted and fused to obtain the joint prediction results for multiple targets or multiple time periods; Using historical data and multi-source fusion information, train the above neural network model in a rolling window manner and perform cross-validation to reduce the risk of overfitting. During the training process, dynamically adjust the weight matrix , bias vector and dynamic weight coefficients and other learnable parameters to obtain a scheduling prediction model, enabling the model to adaptively cope with the impacts of equipment aging, meteorological changes, and random load fluctuations, etc.; By fusing deep neural networks and time series learning units such as gated recurrent units, the collaborative modeling of short-term and ultra-short-term predictions in the same framework is realized, which can greatly improve the overall prediction accuracy; The adaptive weight and bias update mechanism enables the model to respond quickly to changes in the external environment or system state; compared with traditional single algorithms or simple linear regression, this step adopts the method of multi-channel neural network + adaptive weights to provide stronger tracking ability for complex non-linear working conditions (such as rapid meteorological changes or sudden load increases). Through parallel modeling of long time periods (short-term) and high frequency bands (ultra-short-term), this solution takes into account both the overall situation and fine-grained requirements of operation.

[0035] Step 203, Real-time operation evaluation and online model correction After the scheduling prediction model is put into online operation, obtain the actual measurement values at a high frequency (such as every 5 or 15 minutes) (which may include the current real load, PV output, energy storage state, etc.), and compare with the model output for comparison; Define the following custom error function to dynamically measure the prediction accuracy:

[0036] In the formula: is the target index, , is the number of prediction objects to be evaluated simultaneously (such as load, distributed power generation power, energy storage state of charge, etc.); , where: represents the predicted value of the th target, is the corresponding actual observed value; is the relative time after the current evaluation start time , and the value range is , which can be determined according to business requirements of the size (such as 15 minutes, 1 hour or longer); The derivative error weight (non - negative real number) corresponding to the th prediction target, which is used to balance the relative importance of the error magnitude and the error change rate in the comprehensive evaluation; if , then the change trend of the error is not considered, and only the error value itself is measured. When is large, the penalty / reward for the growth or decay rate of the prediction deviation in the evaluation is emphasized; is the time decay factor (non - negative real number) of the th target, which controls the weight allocation of the earlier error and the subsequent error in the evaluation as increases (i.e., as time goes by). When has a large value, the penalty for the error closer to the current time is stronger: if , it means that the errors in the entire interval are accumulated without discrimination; represents the evaluation duration (positive real number), which can be set according to the business scenario. For example, when the evaluation duration is 5 to 15 minutes, it is used for ultra - short - term evaluation, and when the evaluation duration is several hours, it is used for short - term evaluation; it corresponds to the time - scale requirements of short - term and ultra - short - term predictions in step two; According to the fluctuation of the error function , if the error continuously exceeds the preset error threshold, the online correction program is triggered; the latest prediction input vector collected in real - time and the target prediction value are incrementally incorporated into the model training set, and the model parameters are updated through small - scale iterative training, dynamically adjusting the weight matrix , the bias vector and the dynamic weight coefficient etc. to complete the online correction; If it is detected that there is a major change in the system device status (for example, a certain cooling unit is overloaded or some photovoltaic modules fail), it will also interact with this step through the resource status to re - initialize or accelerate the update of the model weights; The real - time evaluation results will be regularly output as a prediction quality report, including the trend of the error function in the recent period and the comprehensive evaluation of the short - term / ultra - short - term prediction accuracy. With the help of the custom error function , the importance of the short - term prediction accuracy is more flexibly highlighted, meeting the requirements of the virtual power plant for rapid response. Through online correction, the model can adaptively track the changes of the device status and external conditions, avoiding the accuracy decay of the pure offline training model in a non - stationary environment; different from the static measurement methods such as the common mean - square error or mean absolute error, the error function Organically integrate the prediction quality at a near moment with the prediction error at a far moment and assign different priorities.

[0037] Step 3: When the prediction result and the resource status are input, through the multi-objective optimization objective function integrate the constraint objectives, use the improved particle swarm algorithm to perform multi-stage optimization on the scheduling decision vector X, construct a hard power balance constraint based on the rated parameters of the equipment, and generate an optimal scheduling plan using the adaptive weight adjustment mechanism , and generate an equipment execution plan, evaluate the effectiveness of the plan through the decomposition of economic-reliable-environmental protection indicators, and achieve the coupling of multi-objective dynamic trade-off and global-local scheduling strategies; The said Step 3 includes the following contents: Step 301: Definition of multi-objective function and constraint modeling According to the short-term / ultra-short-term prediction results, such as load prediction , distributed power generation prediction and energy storage state-of-charge trend prediction , combined with the schedulable resource status, that is, the resource status , construct a multi-objective optimization objective function to simultaneously measure the economic, reliable and environmental protection effects. Let the time domain be , then define:

[0038] Where: , are the intermediate values of the first and second objective functions respectively; X represents the scheduling decision vector to be solved (such as the scheduling or output command of each energy storage, load device, cold and heat energy station, photovoltaic, etc.); , , are the weight coefficients of the objectives (all non-negative real numbers, and at least one is greater than 0); can represent the operating cost function at the current moment , combined with the prediction results , and market prices, feed-in tariffs, etc. for calculation; is used to measure the deviation degree of the system's power supply / energy supply reliability, such as the difference between the load demand and the actual available capacity, or the penalty term generated by insufficient safety margin; then measures the emission level or clean energy utilization rate; can represent the energy offset of the energy storage charge and discharge operation; It can reflect the peak shaving gap or potential risk caused by the mismatch between the current moment and the predicted demand. , are exponential decay factors related to energy storage scheduling and reliability, and their values are all between 0 and 10. Using the rated operating range of equipment, such as the maximum charge and discharge power of energy storage, the upper and lower limits of the capacity of the cold and heat energy station, etc., as the hard constraints of the scheduling decision vector X, referring to the prediction range and time resolution, power balance and safety margin should be ensured in each discrete scheduling period, such as:

[0039] Wherein: represents the power drawn from (or sent to) the power grid, represents the losses of lines and conversion equipment; is the additional available power after peak shifting of adjustable loads or release of energy storage; environmental protection or emission indicators can also be made into hard constraints or soft constraints, such as carbon emission limits, air pollutant emission limits, etc. When in use, by introducing the exponential decay and integral form, more emphasis is placed on the cumulative performance in the entire time domain when measuring the scheduling effect, and the differential dynamic amplification / penalty of frequent operations of energy storage or load differences is carried out, while taking into account economy, reliability and environmental protection, and it can flexibly balance between the multiple objectives of the virtual power plant; compared with using traditional weighted linear models or simple cost functions, this step designs a comprehensive evaluation function of multi-objective + exponential correction and combines the energy storage and load time series information predicted in the second step to control the scheduling smoothness and stability.

[0040] Step 302, implementation of multi-objective optimization algorithm Using the particle swarm position to represent the scheduling decision vector X, and realizing minimization by defining the fitness as the opposite number of the multi-objective optimization objective function or other means; taking the system state (including predicted load, energy storage margin, price information, etc.) within the scheduling period as the state space, and the scheduling decision vector X as the action space, and using the reward or penalty corresponding to the multi-objective optimization objective function to learn the strategy; according to the equipment topology information (such as the number of cold and heat energy stations, the capacity of each energy storage, the types of adjustable loads, etc.) and the prediction duration and time granularity, adjust and set the initial parameters of the algorithm, such as the particle swarm size, exploration step size or the initial network structure of reinforcement learning, etc.

[0041] Enable the algorithm to search for optimization in the search space that matches the equipment scale and time granularity in a targeted manner, thereby improving the solution efficiency and stability for complex energy management problems.

[0042] During the execution process, to balance short-term (such as 1-24 hour scheduling) and ultra-short-term (such as 15-minute rolling scheduling) requirements, the optimization can be divided into multiple stages: Local stage: Refine and correct critical time periods (peak load periods, PV fluctuation periods, etc.) at the minute level; Macro stage: Generate a global scheduling strategy with an hourly step size; When the improved particle swarm optimization algorithm finds that the carbon emissions or safety margin approaches the upper / lower limit during the iteration process, the weight coefficient or proportion can be adaptively increased to explicitly guide the solution towards a greener and more reliable direction; The multi-objective optimization solution process determines convergence by monitoring the iterative changes of the multi-objective optimization objective function and the satisfaction degree of the feasible solution constraints. If full convergence cannot be achieved within the preset number of iterations or time, the current optimal feasible solution is taken, and finally the optimal scheduling plan is output , including: the start-stop sequence of each load device, load distribution, the charge-discharge power of each energy storage at each time period, the priority utilization rate of clean energy (such as PV) and the interaction strategy with the power grid (grid connection, up / down regulation, etc.); When used, the improved particle swarm optimization algorithm can quickly converge in a high-dimensional search space and has strong adaptability to various complex non-linearities and constraints; the multi-stage solution strategy effectively combines short-term and ultra-short-term requirements, improving the feasibility and fineness of the scheduling strategy at different time scales. An adaptive weight and multi-stage combined solution mechanism is introduced, which can balance economic and environmental protection goals at the macro level and finely adjust energy storage and load at the local time period to cope with high-frequency fluctuations or unpredictability.

[0043] Step 303, Scheduling result parsing and next-step connection Map the output optimal scheduling plan to the actual device level and format the output of the scheduling strategy, where: Energy storage device: The charge-discharge power and SOC drift in each sub-time period to ensure matching with the predicted power consumption / generation situation; Cold and heat energy transfer stations and adjustable loads: Specify the start-stop sequence and operating power to ensure meeting the regional heating / cooling and process requirements; Distributed power sources such as PV: Develop a priority grid connection or self-use strategy according to weather forecasts and electricity price fluctuations; Decompose the performance of each item in the optimal scheduling plan in the objective function, for example: Economic cost indicators (such as total power purchase cost or revenue), reliability indicators (such as maximum load difference, number of default penalties); Environmental protection indicators (such as cumulative carbon emissions or renewable energy utilization rate); Execute the optimal scheduling plan and is dynamically adjusted according to the changes in real-time monitoring data and prediction errors; if it is found during execution that the deviation between the prediction and the actual situation intensifies, online optimization or local rescheduling may be triggered to form a closed-loop collaborative mechanism.

[0044] During use, the multi-objective decomposition of the scheduling results helps decision-makers understand the respective achievement degrees in the economic, reliable, and environmental protection dimensions, providing an effective reference for subsequent closed-loop optimization; at the same time, during the multi-objective optimization process, an additional backup scheduling plan or safety mode plan is also output to cope with severe faults or extreme operating conditions: Step Four. When the optimal scheduling plan is sent to on-site equipment, the on-site energy management system converts the instructions into device-level operation sequences, and uses a priority allocation mechanism to synchronously send control instructions and collect monitoring index vectors , and constructs a performance metric function to evaluate the execution deviation. When the deviation accumulates beyond the performance threshold , a rolling correction process is triggered, and local optimal solutions are generated by re-optimizing using a local time window ; The content of the above Step Four includes the following: Step 401. Scheduling instruction issuance and on-site data feedback According to the optimal scheduling plan , the operating power, start-stop status, and charge-discharge plans of each resource such as energy storage, combined heat and power stations, adjustable loads, and distributed power sources in each time period are extracted respectively, and are converted into specific execution instructions for the equipment; The on-site energy management system (EMS) matches the execution points one by one according to the corresponding resource identifiers; Combined with the virtual power plant cloud platform, a priority order is set for the operation instructions of each type of equipment (such as energy storage devices) (such as energy storage scheduling prior to adjustable load peak shaving operation) to ensure that key resources can give priority to executing scheduling tasks in case of emergencies; Using a high-concurrency message queue or a distributed control channel, the instructions are synchronously sent at the same moment or within the same scheduling cycle. For key equipment that is prone to security risks, a confirmation and response link can be added to ensure execution reliability; After receiving the instructions and completing the execution, the on-site energy management system (EMS) sends the actual operating information of the equipment (such as the current state of charge of the energy storage, the power of the combined heat and power unit, the output of the adjustable load, etc.) back to the cloud platform, which together with the data obtained by the sensors constitutes a comprehensive operation monitoring source; During use, through the cascaded mode of scheduling scheme - device mapping - execution feedback, it avoids the actual control chaos caused by inconsistent device positions or identifications. The two-way communication with the cloud platform and the on-site energy management system (EMS) can ensure the real-time visibility of execution instructions. Usually, the scheduling scheme is often issued in the form of static documents, lacking synchronization and feedback tracking. Here, in the design of high-concurrency channels and priority allocation mechanisms, it ensures a low-latency and highly reliable execution process even under the parallel operation of large-scale distributed devices.

[0045] Step 402, Real-time monitoring and system operation evaluation According to the prediction results (such as load , photovoltaic output , energy storage status ), and the optimal scheduling scheme , define real-time monitoring indicators, including but not limited to: Energy storage utilization: The ratio of the actual power output of the energy storage device in the current period to the planned power; Load tracking degree: The difference between the execution output of the adjustable load and the predicted value; Clean energy utilization rate: The proportion of the power generation of distributed power sources such as current photovoltaic in the total energy supply; Safety margin: The remaining space from the rated capacity or safety threshold of the device; The above indicators are summarized in the form of a monitoring indicator vector for use in the next step for evaluation or scheduling correction; To compare the deviation degree between the execution effect of the optimal scheduling scheme and the current operating state, a performance metric function is introduced. By integrating, it takes into account both real-time deviation and change speed. The formula is as follows:

[0046] In the formula: , represents the difference metric function between the current execution deviation or execution result and the scheduling plan at time . If the scheduling is more in line with the actual state, then tends to be closer to 0; the greater the difference, the increases positively; is the executed scheduling instruction, is the real-time monitoring indicator vector; , a function used to measure the scheduling deviation; In , when , the deviation is amplified superlinearly, making large errors account for a higher proportion in the overall evaluation: when In is the weight factor (non - negative real number) of the derivative term, used to balance the importance of the deviation magnitude and the deviation change rate. If then this term gives extra weight to the extreme sudden increase behavior of the error change; and are both control parameters; In where represents the time decay factor; it can be set to 0, so that the deviations in the entire interval are treated equally, is the evaluation window length; If the performance metric function is above the preset performance threshold or a non - linear jump in the energy storage / load device is detected (such as the power deviating extremely from the plan instantaneously), a warning signal is sent to the cloud platform to prompt the dispatching layer or manual operation and maintenance for further inspection; at the same time, it can trigger the entry into the rolling correction process to update the dispatching strategy in a timely manner and avoid accumulating deviations to subsequent time periods.

[0047] When in use, with the help of the performance metric function the influence amount of the dispatching execution deviation can be accumulated in a short time, and the importance of recent abnormal events is highlighted by exponential decay, which is suitable for the dynamic consideration of the synchronous dispatching of multiple devices; traditional real - time monitoring often uses single - point moments or basic statistical indicators for evaluation. Here, the deviations in the entire time period can be continuously observed, enhancing the adaptability to frequently fluctuating scenarios.

[0048] Step 403, Rolling Correction and Closed - loop Collaborative Optimization When the performance metric function exceeds the performance threshold for several consecutive monitoring cycles, or when major anomalies occur in the load, power generation, and energy storage status, the automatic trigger of the dispatching re - evaluation process integrates the latest data collected in real - time (including the time - series fusion sequence and the resource status , the error function , the online correction data of the model, and the execution deviation of the dispatching plan, etc.) to form a new current - moment state vector ; Use the improved particle swarm algorithm or deep reinforcement learning to quickly re - optimize the dispatching plan within a local time window (which can be defined as the next 1 - 2 hours or shorter) to obtain a new local optimal solution After comparing with the original optimal dispatching plan , if it can significantly reduce the recent operation deviation or reduce costs / emissions, the original dispatching instruction can be locally replaced, and the original plan is still used for the remaining time periods. The locally optimal solution after rolling correction Re-enter the instruction issuing process and accept real-time monitoring and evaluation; In use, with the help of small-scale, rapid local re-optimization, it can significantly improve the response speed to real-time fluctuations and avoid the high computational burden brought by large-scale global optimization. The closed-loop collaboration operates continuously under a unified data and algorithm framework, integrating device identification, prediction, and multi-objective optimization into an automatically iterative and updated mechanism, significantly enhancing the practicality and robustness of the solution. While maintaining the global optimal baseline, it flexibly responds to short-term fluctuations and uncertainties, maximizing the operating benefits of the integrated virtual power plant.

[0049] Step Five: When the monitoring index vector and the time-series fusion sequence are abnormal, construct a fault determination function to identify the device-level fault mode. When the fault determination function exceeds the fault threshold , trigger self-healing control, start redundant devices using a backup scheduling plan and adjust the load distribution strategy. After repair, based on the safety threshold restore the device scheduling authority; The content of the above Step Five includes the following: Step 501: Artificial intelligence fault diagnosis Combined with the time-series fusion sequence (such as the operation signals collected from thermal and cold energy stations, adjustable load devices, energy storage, and photovoltaic), and the monitoring index vector (such as the state of charge of energy storage, load tracking degree, etc.; these data are uniformly recorded as comprehensive feedback data , this label covers multi-dimensional information such as high-concurrency acquisition values of sensors, device operation history, and scheduling execution deviation; Introduce a model combining a fault tree (FaultTree) and time-series deep learning (such as LSTM, GRU, or a multi-layer CNN-RNN hybrid structure) to identify potential fault signs at the device and system levels online. The following fault determination function can be defined :

[0050] Where: represents the measurement value of whether the system is at risk of failure at time ; is the possible number of fault modes or leaf nodes (obtained by parsing the fault tree), represents the fault mode index; is the weight coefficient related to the fault mode , a non-negative real number, used to balance the harmfulness of different fault types; is the recognition function of the time-series deep learning model for the input data, Represents a set of trainable parameters for a fault mode ; Represents at time the trigger duration or severity factor for the fault mode ; Represents a fault decay factor, a non - negative real number, used to balance the harmfulness of different fault types; When the fault determination function exceeds the preset fault threshold , it is determined that there is a high fault risk currently. The corresponding device is regarded as a faulty device, and a fault diagnosis instruction is issued. If it is determined that a fault may occur at the device level or system level, it is transmitted to the prediction model construction and real - time operation evaluation in the second step and the intelligent scheduling scheme generation based on multi - objective optimization in the third step: Let the scheduling prediction model consider the possibility of the device operating with a fault or shutting down due to a fault, and the scheduling optimization part also correspondingly adds fault constraints or reduces the schedulable range.

[0051] Trigger the standby scheduling scheme through the monitoring mechanism to prevent the device from continuing to operate in a high - risk state; The standby scheduling scheme includes, specifically: Start of standby equipment: Such as enabling a standby energy storage system, starting additional power generation equipment, or adjusting the load - side response to make up for the loss of the faulty device; Load regulation: In some cases, the load scheduling may need to be adjusted to ensure that the load demand of the system can be met, especially when the faulty device cannot work properly; Adaptive adjustment of the scheduling strategy: During the generation of the standby scheduling scheme, the scheduling strategy will be dynamically adjusted according to information such as fault warnings and prediction deviations to reduce the impact of faults on the overall system performance.

[0052] When in use, with the help of a fault tree, the interpretability of the fault mechanism can be ensured, and temporal deep learning further enhances the detection ability for multi - dimensional sensing data and hidden faults. Traditional fault detection often only relies on static thresholds or simple statistical features. Here, a comprehensive modeling method of fault tree + temporal deep learning + exponential decay factor is introduced, which can not only handle complex non - linear faults, but also dynamically quantify the degree of fault evolution and seamlessly integrate with the multi - objective scheduling system.

[0053] Step 502, Self - healing control and standby scheduling scheme When the fault determination function exceeds the preset fault threshold and shows a rapid upward trend, the system will immediately enter the self - healing control process; Re - evaluate the faulty device or associated load, and transmit instructions to reduce or prohibit the scheduling of the faulty device, or increase the corresponding safety factor; Suspend or terminate the issuance of scheduling instructions for the faulty device, and call the following standby scheduling scheme; After receiving the fault diagnosis instruction, immediately switch to this backup strategy to quickly adapt to the changes in the load and energy supply structure after the fault. Secondary optimization can be performed on non-faulty devices within a local time window to generate a temporary scheduling plan to ensure the supply of the system's core load and the lowest emission requirements. After the self-healing process is started, maintain or isolate the faulty device for a period of time, and gradually observe whether the risk value given by the fault tree model drops significantly. If , where: is the safety threshold; that is, it means that the device fault has been eliminated or the risk is controllable. Then, it can return to the normal scheduling plan, and synchronize the data during the fault period to the data layer in the first step to improve the fault sample library. If the device repair is completed, it can be included in the second-step prediction and the third-step scheduling scope again: if the repair is not completed, it remains in the isolated state to avoid affecting the operation of the overall system.

[0054] When in use, through the pre-deployed backup scheduling plan, as well as the immediate isolation of faulty devices and the secondary optimization of adjacent devices in self-healing control, the risk of fault spread and high-cost power outages is greatly reduced, ensuring the continuous power supply of the system's critical load; after the system returns to normal, the fault diagnosis information and self-healing execution records can further feed back to the data fusion in the first step and the scheduling prediction model in the second step to form a fault history library, improving the accuracy and timeliness of the next fault response; traditional solutions mostly use manual troubleshooting or mechanical shutdowns to handle faults. Here, through the automated mechanism of self-healing control + backup scheduling pre-burial, combined with real-time fault diagnosis determination, not only is the response rapid, but also the scheduling plan can be reconstructed within a short time, with highly intelligent fault control capabilities.

[0055] Those of ordinary skill in the art can realize that the units and algorithm steps of each example described in combination with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are executed in hardware or software depends on the specific application and design constraints of the technical solution. Professional technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of this application.

[0056] Those skilled in the art can clearly understand that for the convenience and brevity of description, the specific working processes of the systems, devices, and units described above can refer to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0057] In several embodiments provided in the present application, it should be understood that the disclosed systems, devices, and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of the units is only for some logical function divisions. In actual implementation, there may be other division methods. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the displayed or discussed coupling or direct coupling or communication connection to each other can be through some interfaces. The indirect coupling or communication connection of the devices or units can be in electrical, mechanical, or other forms.

[0058] The units described as separate components may or may not be physically separated. The components displayed as units may or may not be physical units, that is, they may be located in one place or distributed to multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0059] As described above, this is only the specific implementation manner of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art within the technical scope disclosed in the present application can easily think of changes or substitutions, which should all be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. An intelligent dispatching method for a comprehensive virtual power plant, characterized in that: include, Preprocess the sensor data of distributed power equipment, generate a time series fusion sequence after weighted fusion, build a comprehensive feature vector and determine the resource status by combining equipment operating conditions and environmental data; The initial feature vector is decomposed into a prediction input vector at multiple scales, and the prediction result is generated after parallel training. The prediction deviation is dynamically evaluated using the error function, and online parameter correction is triggered when the error exceeds the error threshold. By integrating constraint objectives through multi-objective optimization objective functions, the scheduling decision vector is optimized in multiple stages, power balance hard constraints are constructed based on equipment rated parameters, and the optimal scheduling solution is generated using an adaptive weight adjustment mechanism; When the optimal scheduling plan is sent to the on-site equipment, the instructions are converted into a device-level operation sequence, and the control instructions are sent synchronously using a priority allocation mechanism and monitoring indicator vectors are collected. A performance measurement function is constructed to evaluate the execution deviation. When the accumulated deviation exceeds the performance threshold, the rolling correction process is triggered. When the monitoring indicator vector and the time series fusion sequence are abnormal and the fault judgment function exceeds the fault threshold, the self-healing control is triggered, the redundant equipment is started, the load distribution strategy is adjusted, and the equipment scheduling authority is restored based on the safety threshold.

2. The intelligent dispatching method of the integrated virtual power plant according to claim 1 is characterized by: Sensors are deployed at distributed power equipment, and the raw data reported by each node and the corresponding timestamp are initially aggregated on the edge side; the acquired raw data is filtered and de-redundant, and the pre-processed data after preliminary screening is obtained, and the synchronous collection and unified timing alignment of different nodes are performed; Import the preprocessed data into the central database, obtain the cleaned data after preprocessing, perform weighted fusion on the data of multiple nodes, and generate a unified single-point fusion value at different times: The single-point fusion values ​​are arranged in sequence to form a time series fusion sequence.

3. The intelligent dispatching method of the integrated virtual power plant according to claim 2 is characterized in that: Integrate the time series fusion sequence with equipment condition records and meteorological data to form a comprehensive feature vector; The comprehensive feature vector is analyzed to determine whether the key resources are in a dispatchable or abnormal state, and the dispatchable state and operation mode of each distributed power device are recorded as the resource state.

4. The intelligent dispatching method of the integrated virtual power plant according to claim 3 is characterized by: From the time series fusion sequence and resource status, select key data and combine with external input to build the initial feature vector; Wavelet transform is performed on some time series components in the initial feature vector to extract high-frequency disturbances and low-frequency trends, decompose the original time series into several multi-scale components, and merge the obtained multi-scale components with other key features to form a prediction input vector.

5. The intelligent dispatching method of the integrated virtual power plant according to claim 4 is characterized in that: Short-term prediction is achieved based on deep neural networks, and ultra-short-term prediction is achieved using lightweight time series models, and key prediction quantities within different prediction time periods are output respectively; The neural network model is trained and cross-validated in a rolling window manner, and the scheduling prediction model is obtained after dynamically adjusting the learnable parameters.

6. The intelligent dispatching method of the integrated virtual power plant according to claim 5 is characterized by: When the scheduling prediction model is put into online operation, the actual measurement values ​​are obtained at a high frequency and compared with the model output, and an error function is defined to dynamically measure the prediction accuracy. According to the fluctuation of the error function, if the error continues to be higher than the preset error threshold, the online correction program is triggered; The latest prediction input vector and target prediction value collected in real time are incrementally incorporated into the model training set, and the model parameters are updated through small-scale iterative training to complete online correction.

7. The intelligent dispatching method of the integrated virtual power plant according to claim 6 is characterized by: According to the short-term / ultra-short-term forecast results of key forecast quantities and combined with resource status, a multi-objective optimization objective function is constructed; Based on the hard constraints of the scheduling decision vector within the rated operating range of the equipment, and referring to the prediction range and time resolution, each discrete scheduling period should ensure power balance and safety margin.

8. The intelligent dispatching method of the integrated virtual power plant according to claim 7 is characterized by: The system state in the scheduling period is used as the state space, the scheduling decision vector is used as the action space, and the reward or penalty corresponding to the multi-objective optimization objective function is used to learn the strategy; Convergence is determined by monitoring the iterative changes of the multi-objective optimization objective function and the satisfaction of the feasible solution constraints. If it cannot be fully converged within the preset number of iterations or time, the current optimal feasible solution is taken and the optimal scheduling plan is finally output; The output optimal scheduling plan is mapped to the actual equipment level, the scheduling strategy is formatted and output, and the performance of the optimal scheduling plan in the objective function is decomposed.

9. The intelligent dispatching method of the integrated virtual power plant according to claim 8 is characterized in that: According to the optimal dispatching plan, the operating power, start-stop status and charging and discharging plan of distributed power equipment in each period are extracted respectively, and converted into specific execution instructions for the equipment; The on-site energy management system matches the execution points one by one according to the corresponding resource identifiers, and combines with the virtual power plant cloud platform to set a priority order for the operating instructions of each type of equipment, and issue the instructions synchronously at the same time or in the same scheduling cycle. After receiving the instructions and completing the execution, the on-site energy management system transmits the actual operation information of the equipment back to the cloud platform, which together with the data obtained by the sensors constitutes a comprehensive operation monitoring source.

10. The intelligent dispatching method of the integrated virtual power plant according to claim 9, characterized in that: The monitoring indicator vector is defined according to the prediction results. In order to compare the deviation between the execution effect of the optimal scheduling plan and the current operating status, when the introduced performance measurement function exceeds the performance threshold for several consecutive monitoring cycles, the scheduling re-evaluation process is automatically triggered, and the latest data collected in real time is integrated to form the state vector at the current moment.

11. The intelligent dispatching method of the integrated virtual power plant according to claim 10, characterized in that: The improved particle swarm algorithm is used to quickly re-optimize the scheduling plan within the local time window. After comparing the new local optimal solution with the original optimal scheduling plan, if it can reduce the recent operating deviation or reduce the cost / emission, the original scheduling instruction can be partially replaced, and the original plan will be used for the rest of the time. The local optimal solution after rolling correction will enter the instruction issuance process again and be subject to real-time monitoring and evaluation.

12. The intelligent dispatching method of the integrated virtual power plant according to claim 11, characterized in that: After combining the time series fusion sequence and the monitoring index vector, they are uniformly recorded as comprehensive feedback data and the fault judgment function is defined; When the fault judgment function exceeds the preset fault threshold, it is determined that there is a high fault risk at present, the corresponding device is regarded as a faulty device, and a fault diagnosis instruction is issued; When the fault judgment function exceeds the preset fault threshold or shows a rapid upward trend, the self-healing control process is immediately entered and the backup scheduling plan is called.

13. The intelligent dispatching method of the integrated virtual power plant according to claim 12, characterized in that: After receiving the fault diagnosis instruction, the backup strategy is immediately switched to quickly adapt to the changes in load and energy supply structure after the fault, and the non-faulty equipment is optimized twice in the local time window to generate a temporary scheduling plan; After the self-healing process is started, the faulty equipment is maintained or isolated, and the risk value given by the fault tree model is gradually observed to see whether it drops. If it drops, it means that the equipment fault has been eliminated or the risk is controllable, and the normal scheduling plan can be returned; If the equipment is repaired, it can be added to the scheduling scope again; if the repair is not completed, it will remain isolated to avoid affecting the overall system operation.

Images