CN-122026510-A - Real-time data processing and feedback adjusting method for household optical storage system

CN122026510ACN 122026510 ACN122026510 ACN 122026510ACN-122026510-A

Abstract

The invention relates to the technical field of digital twin and intelligent control, and discloses a real-time data processing and feedback adjusting method of a household optical storage system. The method comprises the steps of synchronously collecting photovoltaic, energy storage, load and meteorological data, performing timestamp alignment, abnormal rejection and missing interpolation, extracting key features in a lightweight mode on the edge side, uploading the key features to a cloud to construct a mixed digital twin model of physical-data fusion, predicting energy supply and demand in a rolling mode and generating an optimal scheduling instruction, sending down instructions to adjust equipment power, and updating the model on line based on actual response deviation to form closed loop feedback. According to the invention, through the edge cloud cooperative architecture, high-precision and low-delay dynamic adjustment is realized while communication and calculation expenditure is reduced, and the economy, reliability and adaptability of the system are improved.

Inventors

WANG XUEJUN
ZHI HUALIN
HE BAOXIN
MA CHUAN
ZHANG BAOZHONG
WEI JUNYANG
CHEN CHENG

Assignees

河南平煤神马储能有限公司

Dates

Publication Date: 20260512
Application Date: 20251231

Claims (10)

1. The real-time data processing and feedback adjusting method for the household optical storage system is characterized by comprising the following steps of: The method comprises the steps that photovoltaic output data, energy storage battery state data, user load electricity consumption data and local meteorological environment data of a user optical storage system are synchronously obtained through a distributed data acquisition unit; performing time stamp alignment, outlier rejection and missing value interpolation on the acquired multi-source original data to form a standardized time sequence data stream; Performing light feature extraction on the standardized time sequence data stream at an edge computing node to generate an edge feature vector containing power fluctuation rate, charge state change gradient, load peak-valley feature and irradiation intensity trend; Uploading the edge feature vector to a cloud collaborative modeling platform, and constructing a hybrid digital twin model based on physical constraint and data driving fusion on the cloud; Rolling prediction is carried out on the energy supply and demand state in a future preset time window of the system by utilizing the mixed digital twin model, and an optimal charge and discharge scheduling instruction is generated; The optimal charge and discharge scheduling instruction is issued to an edge control unit, and the edge control unit executes real-time power adjustment on the energy storage converter and the photovoltaic inverter; After each adjustment is executed, the actual system response data is collected, deviation analysis is carried out on the actual system response data and the model prediction result, and the parameter weight of the mixed digital twin model is updated on line by utilizing the deviation data, so that a closed-loop feedback adjustment mechanism is formed.
2. The method for real-time data processing and feedback adjustment of a household optical storage system according to claim 1, wherein the distributed data acquisition unit comprises a photovoltaic direct-current side voltage and current sensor, an energy storage battery management system communication interface, an intelligent ammeter data reading module and a miniature weather station, wherein the photovoltaic direct-current side voltage and current sensor acquires output voltage and current of a photovoltaic component, the energy storage battery management system communication interface reads battery cell voltage, temperature, state of charge and health through a controller local area network bus, the intelligent ammeter data reading module acquires user total incoming line and shunt load power through power line carrier communication, and the miniature weather station acquires solar irradiance, ambient temperature, relative humidity and wind speed.
3. The method for real-time data processing and feedback adjustment of a household optical storage system according to claim 1, wherein the time stamp alignment adopts a global synchronization mechanism based on a hardware trigger signal, all data acquisition devices share the same real-time clock source, and generate a synchronization pulse at each whole second moment, and each sensor latches a current sampling value according to the pulse and attaches a uniform time tag; The abnormal value elimination adopts a sliding window 3 times standard deviation criterion, standard deviation and mean value of past data are calculated independently for each data channel, and data points exceeding the standard deviation range of the mean value plus or minus 3 times are marked as abnormal and eliminated; The missing value interpolation adopts a piecewise linear interpolation method, and for a data segment with continuous missing less than 5 sampling points, the data segment is subjected to linear fitting filling by the effective data points before and after the data segment, and if the missing value interpolation is greater than the threshold value, the data segment is marked as an invalid period and the data quality alarm is triggered.
4. The method for real-time data processing and feedback adjustment of a household optical storage system according to claim 1, wherein the lightweight feature extraction deploys a special neural network reasoning engine on an edge computing node, the special neural network reasoning engine loads a feedforward neural network model comprising three full-connection layers, the dimension of an input layer corresponds to a normalized raw data window, the number of neurons of a hidden layer is 64 and 32 respectively, the dimension of an output layer is 16 respectively, and the dimension of the hidden layer corresponds to 16 predefined edge features; The neural network model is subjected to offline training before deployment, a training data set covers historical operation data under different seasons, weather types and user power consumption modes, a loss function adopts a weighted sum of a mean square error and a characteristic physical consistency constraint term, wherein the physical consistency constraint term ensures that the power fluctuation rate is non-negative, and the absolute value of a state of charge change gradient is smaller than a theoretical upper limit corresponding to the maximum charge-discharge multiplying power.
5. The method of claim 4, wherein the 16 predefined edge characteristics include photovoltaic power fluctuation rate, energy storage state of charge gradient, user load peak-to-valley ratio, daily average irradiation intensity trend slope, night load baseline stability index, grid interaction power direction entropy, battery temperature rise rate, photovoltaic module efficiency attenuation factor, load harmonic distortion rate, meteorological mutation sensitivity, energy storage circulation depth cumulative value, photovoltaic-load matching degree, grid electricity price response sensitivity, key load power supply continuity index, energy storage health state attenuation rate and system comprehensive energy efficiency ratio.
6. The method for real-time data processing and feedback adjustment of a household optical storage system according to claim 1, wherein the hybrid digital twin model is formed by connecting a physical mechanism sub-model and a data driving sub-model in parallel; The physical mechanism sub-model is established based on a photovoltaic module equivalent circuit equation, a lithium ion battery electrochemical impedance spectrum simplified model, a load thermodynamic inertia equation and an atmospheric radiation transmission equation, and is input into an original physical quantity in an edge feature vector and output into a theoretical predicted value of a system state variable; The data driving submodel adopts a long-short-period memory network structure and comprises two hidden layers, wherein the input is an edge feature vector, the output is a compensation quantity for a prediction residual error of the physical mechanism submodel, and the final output of the hybrid digital twin model is the sum of the output of the physical mechanism submodel and the output of the data driving submodel.
7. The method for real-time data processing and feedback adjustment of a household optical storage system according to claim 1, wherein the time window length of the rolling prediction is 15 minutes, the prediction step length is 1 minute, i.e. the model generates a power and state of charge sequence of 15 time steps in future per minute; And generating the optimal charge and discharge scheduling instruction by adopting a model prediction control framework, solving a quadratic programming problem with constraint in each control period, wherein an objective function is to minimize the weighted sum of the electricity purchasing cost of a user, the energy storage circulation loss cost and a prediction deviation penalty term, and constraint conditions comprise the upper limit and the lower limit of the energy storage charge state, the power limit of a converter, the interactive power limit of a power grid and the power supply priority requirement of a key load of the user.
8. The method for real-time data processing and feedback adjustment of a household optical storage system according to claim 1, wherein after receiving an optimal charge-discharge scheduling instruction, the edge control unit performs instruction validity check, including whether a power value is within a rated range of the device, whether a state of charge meets a safety boundary, and whether an instruction time stamp is valid; after verification is passed, decomposing the power instruction into an active power set value aiming at the energy storage converter and a maximum power point tracking offset aiming at the photovoltaic inverter, and respectively controlling the energy storage converter and the photovoltaic inverter to execute power adjustment, wherein the end-to-end delay is less than 2 seconds.
9. The method for real-time data processing and feedback adjustment of a household optical storage system according to claim 1, wherein the system response data comprises actual photovoltaic output, battery state of charge, load power and grid interaction power within 5 minutes after the scheduling command is executed, and the sampling frequency is 1 time per second; Calculating the root mean square error of the model predicted value and the actual measured value within 5 minutes of the dispatch instruction execution; And if the root mean square error is greater than the preset threshold value by 5% in continuous 3 control periods, triggering an online model updating process, wherein online updating adopts a migration learning strategy, freezing all parameters of a physical mechanism submodel in the mixed digital twin model, only applying small step gradient descent correction to a weight matrix of a data driving submodel, and using effective operation data of 30 minutes recently as a training sample for each updating.
10. The method for real-time data processing and feedback adjustment of a consumer optical storage system as claimed in claim 9, wherein the root mean square error The calculation formula of (2) is as follows: ; Wherein the method comprises the steps of For the number of sample points, Is the first The predicted value of the model of each predicted step, Is the first Actual measurements are predicted.

Description

Real-time data processing and feedback adjusting method for household optical storage system Technical Field The invention belongs to the technical field of digital twin and intelligent control, and particularly relates to a real-time data processing and feedback adjusting method of a household optical storage system. Background With the rapid popularization of the distributed energy system in a home scene, the household light storage system is used as a key carrier for integrating photovoltaic power generation, electrochemical energy storage and local load management, and the operation efficiency and the regulation precision of the system are highly dependent on the high-efficiency processing capability of multi-source real-time data. The system needs to synchronously collect multidimensional heterogeneous data from a photovoltaic module, an energy storage battery, an electric load, a meteorological environment and the like, and covers various modes such as electric quantity, state quantity, image information, time sequence and the like. However, the conventional data processing architecture generally adopts a traditional data acquisition scheme based on a fixed protocol stack, relies on high-cost hardware and a centralized modeling flow, and is difficult to cope with real challenges such as complicated equipment brands, non-uniform communication protocols, large data quality fluctuation and the like in a home scene. Real-time feedback regulation of the household optical storage system is needed to construct a high-fidelity low-delay dynamic model so as to support core functions such as power scheduling, SOC estimation, fault early warning and the like. The current mainstream method generally realizes modeling through offline training of a static model or deployment of a heavy edge computing unit, wherein the model is misaligned due to incapability of capturing dynamic factors such as equipment aging, weather mutation and the like, and the model is difficult to stably run on a low-cost user terminal for a long time due to calculation force and power consumption limitation. Especially under transient working conditions such as severe illumination change or sudden load increase, the contradiction between the model fidelity and the instantaneity is particularly prominent, and the response speed and the energy efficiency optimization level of the system are severely restricted. In the prior art, when multi-source heterogeneous data are processed, the problems of data island, difficult characteristic alignment, computational redundancy and the like are commonly existed. The traditional fusion method such as weighted average or Kalman filtering is difficult to effectively describe nonlinear coupling relation, while the deep learning model has strong expression capability, but is not suitable for user edge equipment with limited resources due to huge parameter quantity and long training period. In addition, the lack of a lightweight modeling mechanism for data space-time consistency results in a system adjustment strategy that is prone to false positives in the event of low sampling rates or partial sensor failure. Therefore, the realization of high-fidelity dynamic characterization while the light weight of the model is ensured becomes a technical problem to be solved in order to improve the intelligent level of the optical storage system for the user. Disclosure of Invention The invention provides a real-time data processing and feedback adjusting method of a household optical storage system, which aims to solve the technical problems that the traditional data acquisition and modeling mode has high cost and long period and the model fidelity and real-time performance are difficult to achieve. According to the method, a unified access and dynamic cleaning mechanism of multi-source heterogeneous data is constructed, a lightweight edge feature extraction and cloud collaborative modeling architecture is combined, high-fidelity and low-delay sensing of the running state of the user optical storage system is achieved, and the running strategy of the system is dynamically regulated based on a closed-loop feedback mechanism, so that the model accuracy is guaranteed, and meanwhile, the real-time control requirement is met. The invention provides a real-time data processing and feedback adjusting method of a household optical storage system, which comprises the following steps: The method comprises the steps that photovoltaic output data, energy storage battery state data, user load electricity consumption data and local meteorological environment data of a user optical storage system are synchronously obtained through a distributed data acquisition unit; performing time stamp alignment, outlier rejection and missing value interpolation on the acquired multi-source original data to form a standardized time sequence data stream; Performing light feature extraction on the standardized time sequence data stream at an edge c