Distributed energy storage scheduling prediction method and system driven by intelligent environmental monitoring data

Abstract

The application relates to the technical field of energy storage scheduling, and discloses an intelligent environment monitoring data driven distributed energy storage scheduling prediction method and system. The method comprises the following steps: collecting environment parameters to obtain an environment data set; decomposing the environment data in multiple levels to obtain a feature vector and an uncertainty index; inputting a demand prediction distribution of a hybrid deep network; constructing a robust optimization model to obtain a scheduling instruction sequence; decomposing and executing a running track through a hierarchical architecture; and iteratively optimizing an energy storage scheduling strategy through deep reinforcement learning. The application overcomes the influence of communication delay and has a self-optimization capability of an energy storage scheduling prediction method, and realizes efficient collaborative operation of a distributed energy storage system based on environment data driving.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of energy storage scheduling, and particularly relates to an intelligent environment monitoring data driven distributed energy storage scheduling prediction method and system. BACKGROUND

[0002] With the rapid development of renewable energy and the promotion of power system transformation, energy storage technology as an important support means has attracted widespread attention. In the traditional power system, the coordination of power generation, transmission, distribution and power consumption relies mainly on the regulation ability of large power plants to maintain the balance between supply and demand. However, the intermittency and volatility of renewable energy sources such as wind and solar energy have brought new challenges to the stable operation of the power system. In view of this problem, distributed energy storage systems have become a key technology for improving renewable energy consumption capacity and enhancing power grid stability due to their flexibility and rapid response capability. At present, common energy storage scheduling methods mainly include statistical prediction based on historical data, deterministic optimization based on physical models and intelligent algorithms based on machine learning. These methods predict power consumption load and renewable energy generation capacity, combine electricity price information and grid constraints, and develop energy storage charging and discharging strategies to achieve a balance between economy and reliability.

[0003] However, the existing energy storage scheduling methods have obvious shortcomings. First, traditional deterministic optimization methods are difficult to effectively handle the uncertainty brought by environmental factors, especially in the case of large renewable energy prediction errors, the scheduling results are often too idealistic, and the actual execution effect is not satisfactory. Second, most methods lack sufficient use of environmental monitoring data, and fail to establish an accurate mapping relationship between environmental factors and energy storage demand, resulting in limited prediction accuracy. Third, centralized scheduling strategies have high computational complexity and poor real-time performance when facing large-scale distributed energy storage systems, and are difficult to adapt to rapidly changing environmental conditions. Fourth, existing methods usually adopt deterministic control execution mode, and lack effective response mechanism to actual problems such as communication delay, making it difficult to guarantee control accuracy and system stability. Finally, there is a lack of adaptive optimization mechanism, which cannot continuously improve the scheduling strategy according to the actual operation results, making it difficult to achieve long-term performance optimization. SUMMARY

[0004] The present application provides an intelligent environment monitoring data driven distributed energy storage scheduling prediction method and system, which overcomes the communication delay and has self-optimization ability of energy storage scheduling prediction method, realizes efficient collaborative operation of distributed energy storage system based on environmental data.

[0005] In a first aspect, the application provides an intelligent environment monitoring data driven distributed energy storage scheduling prediction method, which comprises: processing temperature, humidity, wind speed, light intensity, and atmospheric pressure data collected by a distributed environment monitoring node to obtain an environment data set; performing multi-level decomposition on the environment data set to obtain an environment feature vector and an uncertainty quantization index; inputting the environment feature vector and the uncertainty quantization index into a hybrid deep learning model composed of a multi-layer BiLSTM network and a time convolution network to obtain an energy storage demand prediction distribution; constructing an energy storage scheduling model with a two-layer optimization structure according to the energy storage demand prediction distribution, converting a distribution robust optimization problem into a mixed integer quadratic programming, and obtaining an energy storage scheduling instruction sequence; decomposing the energy storage scheduling instruction sequence through a hierarchical distributed architecture, distributing global optimization instructions to each energy storage unit, executing through a non-periodic sampling data driven model predictive controller, and obtaining energy storage unit operation trajectory data; and iteratively optimizing the energy storage unit operation trajectory data through a deep deterministic policy gradient algorithm to obtain an energy storage scheduling strategy.

[0006] In a second aspect, the application provides an intelligent environment monitoring data driven distributed energy storage scheduling prediction system, which comprises:

[0007] A processing module for processing temperature, humidity, wind speed, light intensity, and atmospheric pressure data collected by a distributed environment monitoring node to obtain an environment data set;

[0008] A decomposition module for performing multi-level decomposition on the environment data set to obtain an environment feature vector and an uncertainty quantization index;

[0009] An input module for inputting the environment feature vector and the uncertainty quantization index into a hybrid deep learning model composed of a multi-layer BiLSTM network and a time convolution network to obtain an energy storage demand prediction distribution;

[0010] A conversion module for constructing an energy storage scheduling model with a two-layer optimization structure according to the energy storage demand prediction distribution, converting a distribution robust optimization problem into a mixed integer quadratic programming, and obtaining an energy storage scheduling instruction sequence;

[0011] A decomposition module for decomposing the energy storage scheduling instruction sequence through a hierarchical distributed architecture, distributing global optimization instructions to each energy storage unit, executing through a non-periodic sampling data driven model predictive controller, and obtaining energy storage unit operation trajectory data;

[0012] An optimization module for iteratively optimizing the energy storage unit operation trajectory data through a deep deterministic policy gradient algorithm to obtain an energy storage scheduling strategy.

[0013] In a third aspect, an intelligent environment monitoring data-driven distributed energy storage scheduling prediction device is provided, comprising a memory and at least one processor, wherein the memory stores instructions; and the at least one processor invokes the instructions in the memory to enable the intelligent environment monitoring data-driven distributed energy storage scheduling prediction device to perform the intelligent environment monitoring data-driven distributed energy storage scheduling prediction method described above.

[0014] In a fourth aspect, a computer-readable storage medium is provided, wherein the computer-readable storage medium stores instructions that, when executed on a computer, enable the computer to perform the intelligent environment monitoring data-driven distributed energy storage scheduling prediction method described above.

[0015] In the technical scheme provided in the present application, the temperature, humidity, wind speed, light intensity and atmospheric pressure data collected by the distributed environment monitoring nodes are processed to establish a direct mapping relationship between the environment data and the energy storage demand, thereby improving the data utilization efficiency and making the energy storage scheduling more in line with the actual environmental conditions. The environment data set is decomposed at multiple levels to obtain an environment feature vector and an uncertainty quantization index, thereby realizing the extraction of deep-level features of the environment data and the accurate quantization of uncertainty. The environment feature vector and the uncertainty quantization index are input into a hybrid deep learning model composed of a multi-layer BiLSTM network and a time convolution network, thereby fully utilizing the advantages of BiLSTM in capturing long-term dependencies and TCN in processing long sequence data, and making the energy storage demand prediction distribution more accurate and reliable. In particular, when applying artificial intelligence algorithms, the network structure design of the hybrid deep learning model takes into account the time sequence characteristics and multi-dimensional features of the environment data, and the feature weights are adaptively allocated through the attention mechanism, thereby significantly improving the prediction accuracy. According to the energy storage demand prediction distribution, an energy storage scheduling model with a two-layer optimization structure is constructed, and a distribution robust optimization problem is converted into a mixed integer quadratic programming, thereby effectively handling the influence of environmental uncertainty on scheduling and enhancing the robustness of the scheduling strategy. The energy storage scheduling instruction sequence is decomposed through a hierarchical distributed architecture, and the global optimization instruction is distributed to each energy storage unit, which is executed by a non-periodic sampling data-driven model predictive controller, thereby solving the problems of computational complexity and communication delay of traditional centralized control in large-scale distributed systems, and improving the control accuracy and system response speed. The energy storage unit operation trajectory data is iteratively optimized through a deep deterministic policy gradient algorithm, thereby realizing adaptive optimization and continuous improvement of the scheduling strategy, and continuously improving the system performance over time. In the application of artificial intelligence algorithms, the introduction of deep reinforcement learning technology enables the system to learn and improve the strategy from actual operation experience, which is particularly suitable for handling complex problems involving sequential decision-making such as energy storage scheduling. The whole process of environment monitoring, demand prediction, scheduling optimization, collaborative control and strategy optimization is realized. BRIEF DESCRIPTION OF DRAWINGS

[0016] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed to be used in the embodiments will be briefly introduced as follows. Obviously, the drawings in the following description are some embodiments of the present application, and other drawings can be obtained by those skilled in the art without any creative effort based on these drawings.

[0017] Figure 1 An embodiment schematic diagram of the intelligent environment monitoring data driven distributed energy storage scheduling prediction method in the embodiments of the present application;

[0018] Figure 2 An embodiment schematic diagram of the intelligent environment monitoring data driven distributed energy storage scheduling prediction system in the embodiments of the present application;

[0019] Figure 3 An embodiment schematic diagram of the intelligent environment monitoring data driven distributed energy storage scheduling prediction system in the embodiments of the present application; DETAILED DESCRIPTION

[0020] The embodiments of the present application provide an intelligent environment monitoring data driven distributed energy storage scheduling prediction method and system. The terms "first", "second", "third", "fourth" and the like (if any) in the specification and claims of the present application and the above drawings are used to distinguish similar objects, and do not necessarily have to be used to describe a specific order or sequence. It should be understood that the data thus used can be interchanged under appropriate circumstances, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the term "comprising" or "having" and any variation thereof is intended to cover non-exclusive inclusion, for example, a process, method, system, product or device including a series of steps or units does not have to be limited to those steps or units clearly listed, but can include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0021] For the convenience of understanding, the specific process of the embodiments of the present application will be described below. Please refer to Figure 1 An embodiment of the intelligent environment monitoring data driven distributed energy storage scheduling prediction method in the embodiments of the present application includes:

[0022] Step S101, processing the temperature, humidity, wind speed, light intensity, atmospheric pressure data collected by the distributed environment monitoring node to obtain an environment data set;

[0023] Step S102, performing multi-level decomposition on the environment data set to obtain an environment feature vector and an uncertainty quantization index;

[0024] Step S103, input the environmental feature vector and the uncertainty quantization index into a hybrid deep learning model composed of a multi-layer BiLSTM network and a time convolution network, to obtain a energy storage demand prediction distribution;

[0025] Step S104, constructing an energy storage scheduling model with a two-layer optimization structure according to the energy storage demand prediction distribution, converting the distribution robust optimization problem into a mixed integer quadratic programming, to obtain an energy storage scheduling instruction sequence;

[0026] Step S105, decomposing the energy storage scheduling instruction sequence through a hierarchical distributed architecture, distributing the global optimization instruction to each energy storage unit, and executing through a non-periodic sampling data driven model predictive controller, to obtain energy storage unit operation trajectory data;

[0027] Step S106, iteratively optimizing the energy storage unit operation trajectory data through a deep deterministic policy gradient algorithm, to obtain an energy storage scheduling strategy.

[0028] It can be understood that the execution subject of the present application can be an intelligent environment monitoring data driven distributed energy storage scheduling prediction system, and can also be a terminal or a server, which is not limited here. The server is taken as an example for illustration in the embodiments of the present application.

[0029] Specifically, the process of processing the temperature, humidity, wind speed, light intensity, and atmospheric pressure data collected by the distributed environment monitoring nodes to obtain the environmental data set actually includes multiple data processing operations. In this step, the distributed environment monitoring nodes are arranged around the energy storage units to form a monitoring grid, and the environmental data is collected every 10 minutes. The collected raw data is transmitted to the edge computing node through the wireless sensor network, and then wavelet transform denoising processing is performed to remove high-frequency noise interference. Next, data normalization is performed to unify the environmental parameters of different dimensions to the [0, 1] interval, which is convenient for subsequent processing. For data missing points, an interpolation algorithm based on time series correlation is used for repair to ensure the continuity of the data. At the same time, the 3σ principle and the density clustering algorithm are combined to identify outliers and mark them appropriately. Finally, the processed data is reorganized into a multi-dimensional time series matrix in chronological order, and the statistical characteristics are calculated to form a complete environmental data set.

[0030] A multi-level decomposition of the environmental dataset was performed to obtain an environmental feature vector and uncertainty quantification indicators. This step employed a multi-level decomposition strategy, decomposing the environmental data into trend, periodic, and random fluctuation components. The trend component was extracted using a local polynomial fitting method, with a polynomial order of 3 and a sliding window width of 24 hours. The periodic component was extracted using a fast Fourier transform combined with adaptive filtering to identify diurnal and seasonal variations. The random fluctuation component was obtained through a differential operation, representing the short-term random perturbations of the environmental parameters. The characteristics of the three components were then calculated separately and concatenated to form a 128-dimensional environmental feature vector. To quantify environmental uncertainty, a probability distribution family representation framework based on fuzzy set theory was constructed. An empirical distribution function was established for each environmental parameter, and a 0.95 confidence interval was defined to form the upper and lower bounds of the distribution function. Furthermore, an extreme value distribution fitting method was used to account for the extreme value distribution characteristics of the environmental parameters, and a copula function was used to capture the dependencies between the multidimensional environmental parameters.

[0031] The environmental feature vector and uncertainty quantification index are input into a hybrid deep learning model consisting of a multi-layer BiLSTM network and a temporal convolutional network to generate the predicted energy storage demand distribution. This step first uses an attention mechanism to adaptively assign weights to environmental features. The attention weights are calculated as the product of a soft-maximization function, a hyperparameter vector, and the environmental feature vector. Then, through residual connections and layer normalization, a feature fusion vector is obtained. This vector is input into a three-layer BiLSTM network with a hidden dimension of 256. A dropout layer with a dropout rate of 0.2 is added between each layer to capture temporal dependencies. Next, the temporal feature representation enters a temporal convolutional network consisting of eight dilated convolutional layers with dilation factors of 1, 2, 4, 8, 16, 32, 64, and 128, a convolution kernel size of 3, and 64 output channels. This generates the probability distribution parameters for energy storage demand. Based on these parameters, a conditional Gaussian distribution is constructed and Monte Carlo random sampling is performed 100 times in combination with the uncertainty quantification index. After calibration using a quantile regression loss function, the predicted energy storage demand distribution is obtained.

[0032] A two-level optimization structure is built for the energy storage scheduling model according to the energy storage demand prediction distribution, and the distribution robust optimization problem is converted into a mixed integer quadratic programming to obtain the energy storage scheduling instruction sequence. First, a target function is built, which includes the charge and discharge loss cost, the capacity adjustment cost and the response deviation penalty cost, to form a distribution robust optimization main problem that minimizes the total energy storage operation cost in the worst case. The state of charge constraint, the charge and discharge power constraint, the charge and discharge switching constraint and the distribution robust opportunity constraint are set. The Wasserstein distance is introduced to limit the distance between the probability distribution of the prediction error of the environmental parameters and the empirical distribution. Through the auxiliary variable and the strong duality theorem, the whole problem is converted into a mixed integer quadratic programming model. Subsequently, the hierarchical decomposition algorithm is used to decompose the global scheduling problem into multiple sub-regional coordination problems, and the alternating direction multiplier method is used to solve each sub-problem. Finally, the non-active constraint identification technology and the convex relaxation technology are applied to optimize the solution of each sub-region to obtain the energy storage scheduling instruction sequence for the next 24 hours.

[0033] The energy storage scheduling instruction sequence is decomposed through a hierarchical distributed architecture to allocate the global optimization instruction to each energy storage unit, which is executed by a non-periodic sampling data-driven model predictive controller to obtain the energy storage unit operation trajectory data. This process first builds a hierarchical distributed architecture including a coordination layer, an execution layer and a monitoring layer, and sets a directed acyclic graph as the communication topology. The global scheduling instruction is decomposed into local control instructions for each energy storage unit through a consistent convergence distributed algorithm, which is allocated according to the principle of minimizing the global coordination cost. A state space model is established for each energy storage unit, and an appropriate prediction and control time domain is set to obtain the mathematical expression of the model predictive controller. For the communication delay problem, a non-periodic sampling data-driven control method is designed, and the controller gain matrix is obtained by solving the linear matrix inequality. A monitoring module is deployed at the energy storage unit to collect real-time operation parameters and calculate performance indicators. The system stability is verified by constructing the Lyapunov-Krasovskii functional, and the energy storage unit operation trajectory data is finally recorded.

[0034] The trajectory data of energy storage unit operation is iteratively optimized by deep deterministic policy gradient algorithm to obtain the energy storage scheduling strategy. A multi-dimensional evaluation system is first constructed to calculate the comprehensive performance score of the trajectory data and form the feedback evaluation signal. The energy storage scheduling process is modeled as a Markov decision process, and appropriate state space, action space and reward function are defined. A policy network and a value network containing 4 layers of fully connected neural networks are constructed, with the number of nodes in each layer being 512, 256, 128 and 64 respectively, and the activation function being a leaky rectified linear unit. An experience replay buffer is established, and reasonable parameters such as batch size, discount factor, learning rate and target network soft update coefficient are set. Parameter space noise is added to enhance the exploration ability, and the current policy is applied to the actual system every 100 training batches to collect performance data and update the experience replay buffer. Through continuous optimization and adjustment of the parameters of each module, an efficient energy storage scheduling strategy is finally formed.

[0035] In an embodiment of the present application, by processing the temperature, humidity, wind speed, light intensity, and atmospheric pressure data collected by distributed environmental monitoring nodes, a direct mapping relationship between environmental data and energy storage demand is established, which improves data utilization efficiency and makes energy storage scheduling more in line with actual environmental conditions. The environmental data set is decomposed at multiple levels to obtain environmental feature vectors and uncertainty quantification indicators, which realizes the extraction of deep-level features of environmental data and accurate quantification of uncertainty. The environmental feature vectors and uncertainty quantification indicators are input into a hybrid deep learning model composed of a multi-layer BiLSTM network and a time convolutional network, giving full play to the advantages of BiLSTM in capturing long-term dependencies and TCN in processing long sequence data, making the energy storage demand prediction distribution more accurate and reliable. In particular, when applying artificial intelligence algorithms, the network structure design of the hybrid deep learning model takes into account the temporal characteristics and multidimensional characteristics of environmental data, and adaptively allocates feature weights through the attention mechanism, significantly improving the prediction accuracy. According to the predicted distribution of energy storage demand, a two-layer optimization structure energy storage scheduling model is constructed, and the distributed robust optimization problem is converted into a mixed integer quadratic programming, which effectively handles the impact of environmental uncertainty on scheduling and enhances the robustness of the scheduling strategy. The energy storage scheduling instruction sequence is decomposed through a hierarchical distributed architecture, and global optimization instructions are distributed to each energy storage unit. They are executed through a model predictive controller driven by non-periodic sampling data. This solves the computational complexity and communication delay problems of traditional centralized control in large-scale distributed systems, and improves control accuracy and system response speed. The operation trajectory data of the energy storage unit is iteratively optimized through a deep deterministic policy gradient algorithm, achieving adaptive optimization and continuous improvement of the scheduling strategy, so that the system performance continues to improve over time. In terms of the application of artificial intelligence algorithms, the introduction of deep reinforcement learning technology enables the system to learn and improve strategies from actual operating experience, which is particularly suitable for dealing with complex problems such as energy storage scheduling that involve sequential decision-making. The full process of environmental monitoring, demand forecasting, scheduling optimization, collaborative control and strategy optimization has been realized.

[0036] In a specific embodiment, the process of executing step S101 may specifically include the following steps:

[0037] The temperature, humidity, wind speed, light intensity, and atmospheric pressure data collected by distributed environmental monitoring nodes are subjected to high-frequency noise removal using the wavelet transform method to obtain denoised environmental data.

[0038] Normalize the denoised environmental data to unify the numerical ranges of different types of environmental parameters to the interval [0,1] to obtain normalized environmental data;

[0039] Based on the temporal correlation, the missing points in the normalized environmental data are repaired using an interpolation algorithm to obtain complete environmental data;

[0040] The complete environmental data is combined with the 3σ principle and the density clustering algorithm for anomaly detection and labeling to obtain labeled environmental data;

[0041] The labeled environmental data is reorganized into a multi-dimensional time series matrix in chronological order to obtain reconstructed environmental data;

[0042] The reconstructed environmental data is calculated for mean, variance, skewness, kurtosis, and autocorrelation coefficient, and the statistical characteristics are combined to form the basic components of the environmental feature vector, obtaining the environmental data set.

[0043] Specifically, the temperature, humidity, wind speed, light intensity, and atmospheric pressure data collected by the distributed environmental monitoring nodes are processed using the wavelet transform method to remove high-frequency noise and obtain denoised environmental data. The wavelet transform method is suitable for environmental data processing because environmental signals often contain useful information and noise of different frequencies. In specific operations, a wavelet basis function suitable for the characteristics of environmental data (such as Daubechies wavelet, Haar wavelet, etc.) is selected, the original signal is decomposed into wavelet coefficients of different frequency intervals, then the high-frequency coefficients representing noise are thresholded, and finally the signal is reconstructed by wavelet inverse transform. For example, common high-frequency electrical noise and sampling fluctuations in temperature data are effectively filtered out by selecting an appropriate threshold function, while retaining the true trend characteristics of temperature changes. The denoised environmental data is normalized to unify the numerical range of different types of environmental parameters to the [0, 1] interval, obtaining normalized environmental data. Normalization takes into account the significant differences in units and magnitudes of different environmental parameters, such as temperature units in Celsius usually fluctuating between -20 and 50, while atmospheric pressure units in hectopascals usually varying between 900 and 1100, and light intensity possibly ranging from 0 to tens of thousands of lux. The normalization process uses the maximum and minimum value normalization method to calculate the historical maximum and minimum values of each parameter type, then adjusts the data range to the [0, 1] interval according to linear mapping. The processed data facilitates subsequent training and calculation of neural network algorithms.

[0044] The missing points in the normalized environmental data are repaired by interpolation algorithm based on time correlation to obtain complete environmental data. In the actual environmental monitoring process, data missing is a common phenomenon due to sensor failure, communication interruption, etc. In view of this situation, the position of missing points in the data sequence is first identified, and then the appropriate interpolation method is selected according to the time series characteristics of environmental parameters. For short-time missing points, linear interpolation algorithm is adopted; for longer time missing, considering the time correlation, weighted average interpolation is carried out combined with historical data at the same period and adjacent time point data; for environmental parameters with obvious periodicity such as daily variation of temperature, periodic interpolation method is adopted to estimate the missing value by using the data at the same period of the previous days. The complete environmental data is combined with 3σ principle and density clustering algorithm for abnormal value detection and marking to obtain the marked environmental data. 3σ principle is an abnormal detection method based on statistics, which assumes that the data obeys normal distribution, and the data points outside the range of mean ± 3 times standard deviation are regarded as potential anomalies. The density clustering algorithm identifies abnormal values by analyzing the density distribution of data points. Common density clustering algorithms such as DBSCAN (density-based spatial clustering of applications with noise) can effectively identify outliers in environmental data. In specific operation, first calculate the mean and standard deviation of each type of environmental parameter according to the time window segmentation, apply 3σ principle to preliminarily mark suspicious points, and then use density clustering algorithm to further verify in multi-dimensional feature space to finally determine abnormal points. For the marked abnormal values, instead of being directly deleted, a mark is added so that these data can be selectively used or specially processed in subsequent processing.

[0045] Different types of environmental parameters (temperature, humidity, wind speed, etc.) and data from different monitoring nodes are integrated into a structured data matrix according to a unified time axis. The rows of the matrix represent the sampling points of the time series, and the columns represent different environmental parameters and monitoring node positions. In specific implementation, all data are sorted according to the time stamp to ensure time point alignment and form a uniform sampling time series. For parameters with inconsistent monitoring frequencies, interpolation is used to match them to the unified time points. This structured multi-dimensional time series matrix facilitates subsequent analysis of the spatio-temporal correlation between environmental parameters. The mean, variance, skewness, kurtosis, and autocorrelation coefficient of the reconstructed environmental data are calculated to form the basic components of the environmental feature vector, and the environmental data set is obtained. Statistical feature calculation is to extract the distribution characteristics and time correlation of environmental data. The mean reflects the central tendency of the data; the variance represents the dispersion of the data; the skewness measures the asymmetry of the distribution, with positive skewness indicating right-skewed distribution and negative skewness indicating left-skewed distribution; the kurtosis measures the sharpness of the distribution, with high kurtosis indicating a sharp peak and low kurtosis indicating a flat distribution; the autocorrelation coefficient reflects the correlation between the time series data and its own lag version, which helps to capture the periodicity and persistence characteristics of the data. By calculating the autocorrelation coefficient at different time scales, the daily variation pattern and seasonal trend of the environmental parameter can be identified. These statistical features form the basic components of the environmental feature vector, providing a data basis for subsequent multi-level decomposition and uncertainty quantification of environmental data.

[0046] In a specific embodiment, the process of performing step S102 can specifically include the following steps:

[0047] A local polynomial fitting method is used to extract the trend component from the environmental data set, with a polynomial order of 3 and a sliding window technique, to obtain the trend component of the environmental parameter;

[0048] Fast Fourier transform is applied to the residual data obtained by subtracting the trend component from the environmental data set, and adaptive filtering is used to obtain the periodic component of the environmental parameter;

[0049] Difference operation is performed on the data obtained by subtracting the trend component and the periodic component from the environmental data set to obtain the random fluctuation component of the environmental parameter;

[0050] The slope, curvature, and number of inflection points are calculated based on the trend component, the main frequency, amplitude, and phase are calculated based on the periodic component, and the fluctuation range, mutation frequency, and duration are calculated based on the random fluctuation component. The three types of feature vectors are concatenated to obtain

[0051] An empirical distribution function is established based on historical observation data for each environmental parameter in the environmental data set, and a confidence interval with a confidence level of 0.95 is defined to construct the upper and lower bounds of the distribution function, obtaining a distribution family containing all possible probability distributions;

[0052] The Wasserstein distance metric is used based on the distribution family to calculate the upper bound of the distance between the worst-case distribution and the reference distribution, and the Copula function is used to capture the correlation structure between multidimensional environmental parameters to obtain the uncertainty quantification index.

[0053] Specifically, a local polynomial fitting method is used to extract the trend component of the environmental data set. The polynomial order is set to 3, and a sliding window technique is used to obtain the trend component of the environmental parameters. Local polynomial fitting is a nonparametric regression method suitable for extracting long-term trends in environmental data. In practice, a time window (e.g., 24 hours) is selected, within which a third-order polynomial function is fitted to the environmental data. The polynomial coefficients are determined using the least squares method. This window is then slid along the time axis, and the fitted value is calculated for each time point, resulting in a complete trend component curve. Compared to global fitting, local polynomial fitting is more adaptable to the nonlinear variations in environmental data, capturing turning points and gradual changes in the data. Fast Fourier transform (FFT) is applied to the residual data after subtracting the trend component from the environmental data set, combined with adaptive filtering, to obtain the periodic components of the environmental parameters. The Fast Fourier Transform (FFT) is an efficient discrete Fourier transform algorithm that converts time-domain signals into frequency-domain representations, revealing periodic components in the signal. In environmental data processing, the FFT of the residual data is first calculated to generate a spectrogram. The primary frequency components are then identified from the spectrogram. Environmental data typically exhibit a distinct diurnal cycle (with a frequency of 1 / 24 hour) and possible seasonal cycles. An adaptive filter is then applied to retain the identified primary periodic components while filtering out non-periodic noise. The parameters of the adaptive filter are dynamically adjusted based on the data characteristics to ensure accurate extraction of periodic components without introducing spurious periods. An inverse FFT transform is then performed to convert the selected frequency-domain signal back to the time domain, yielding the periodic components of the environmental parameters.

[0054] Differencing the environmental data set after subtracting the trend and cyclical components yields the random fluctuation components of the environmental parameters. Differencing is a commonly used processing method in time series analysis, effectively extracting short-term random fluctuations from the data. Specifically, the difference between data points at adjacent time points is calculated, expressed as x'(t) = x(t) - x(t-1), where x(t) represents the data value at time t and x'(t) represents the differenced value. For environmental data, first-order differencing is generally sufficient to reveal random fluctuations; higher-order differencing can also be considered for data with strong autocorrelation. Differencing can eliminate the effects of residual low-frequency variations and highlight the random disturbances of environmental parameters, such as those caused by sudden weather events and temporary load changes.

[0055] The slope is obtained by calculating the first derivative of the trend curve, reflecting the rate of change of the environmental parameter; the curvature is obtained by calculating the second derivative, representing the bending degree of the trend curve; the number of inflection points is obtained by detecting the number of sign changes of the second derivative, representing the complexity of the trend change. For the periodic component, the main frequency is determined by the peak frequency in the spectrum analysis; the amplitude is the amplitude of the corresponding frequency component, reflecting the strength of the periodic change; the phase represents the time offset of the periodic fluctuation. The fluctuation range of the random fluctuation component is calculated as the difference between the maximum and minimum values; the mutation frequency is obtained by detecting the number of fluctuations exceeding the threshold; the duration is the distribution of the length of continuous fluctuations. These features comprehensively describe the static statistical characteristics and dynamic change rules of the environmental data, and form a complete environmental feature representation through vector concatenation. For each environmental parameter in the environmental data set, an empirical distribution function is established based on historical observation data, and a confidence interval with a confidence level of 0.95 is defined to construct the upper and lower bounds of the distribution function, obtaining a distribution family containing all possible probability distributions. The empirical distribution function (EDF) is a non-parametric statistical method that directly constructs the cumulative distribution function based on observation data without assuming a specific distribution form. In specific implementation, the historical observation values of each environmental parameter are sorted by size, the cumulative probability is calculated, and a step-shaped empirical distribution function is formed. The confidence interval with a confidence level of 0.95 is calculated by the DKW inequality (Dvoretzky-Kiefer-Wolfowitz inequality), which gives the upper bound of the probability that the empirical distribution function deviates from the true distribution function. By constructing the upper and lower bounds of the empirical distribution function, a distribution family is formed, which contains all possible probability distributions consistent with the statistical data of the historical observation data, providing a mathematical basis for the quantification of environmental uncertainty.

[0056] The upper bound of the distance between the worst-case distribution and the reference distribution is calculated based on the distribution family using the Wasserstein distance metric, and the correlation structure between the multi-dimensional environmental parameters is captured using the Copula function to obtain the uncertainty quantification index. The Wasserstein distance is a measure of the difference between two probability distributions, which has good geometric interpretation and computational properties. In environmental uncertainty quantification, the upper bound of the Wasserstein distance between any distribution in the distribution family and the reference distribution (usually the empirical distribution) is calculated to represent the maximum possible deviation of the environmental parameter prediction. The Copula function is a powerful tool for describing the joint distribution of multiple variables, which can "couple" each edge distribution together to form a complete joint distribution. In the processing of multi-dimensional environmental parameters, the edge distribution of each parameter is established first, and then the dependence structure between the parameters is captured through the Copula function (such as Gaussian Copula, t-Copula, Archimedean Copula, etc.). The Copula family that best fits the historical data is selected, and the parameters are determined through maximum likelihood estimation to construct the joint distribution model of the environmental parameters. The uncertainty quantification index calculated based on the Wasserstein distance and the Copula model comprehensively describes the distribution characteristics and correlation structure of the environmental parameters, providing uncertainty characterization for subsequent robust energy storage scheduling.

[0057] Taking a certain distributed new energy microgrid project as an example, the environmental feature extraction process processes one month of environmental monitoring data. When using local polynomial fitting on the light intensity data, a 24-hour sliding window and a 3rd order polynomial are selected, and the daily variation trend of the light intensity is successfully extracted. The average slope of the morning period is 180 lux / hour, the average slope of the afternoon period is-160 lux / hour, and there is an obvious inflection point at sunrise and sunset. The FFT analysis of the residual signal shows obvious daily and weak semi-daily cycles, with amplitudes of about 80% and 15% of the daily average light, respectively. The difference calculation of the random fluctuation component shows that the short-term fluctuation is mainly affected by the change of cloud layer, and the fluctuation range reaches ±25% of the average value, and the typical mutation duration is 15-30 minutes. By constructing the empirical distribution function of the light intensity and the 0.95 confidence interval, a distribution family is formed, and combined with other environmental parameters such as temperature, a multi-dimensional joint distribution model is established using the Gaussian Copula function, successfully capturing the positive correlation between the light intensity and the temperature. The upper bound of the Wasserstein distance calculated shows that in the worst case, the deviation of the environmental parameter prediction is not more than 1.8 times the historical standard deviation. These environmental features and uncertainty quantification indexes are directly input into the subsequent energy storage demand prediction model.

[0058] In a specific embodiment, the process of performing step S103 can specifically include the following steps:

[0059] The attention mechanism is adopted for adaptive weight distribution of the environmental feature vector, and the attention weight value is calculated as the product of a soft-max function and a hyperparameter vector and the environmental feature vector, to obtain a weighted environmental feature vector;

[0060] The weighted environmental feature vector is connected through a residual connection and normalized by a layer to obtain a feature fusion vector;

[0061] The feature fusion vector is input into a bidirectional long short-term memory network with a hidden layer dimension of 256 and 3 layers, and a Dropout layer with a dropout rate of 0.2 is added between each layer to obtain a time sequence feature representation that captures time dependence;

[0062] The time sequence feature representation is input into a temporal convolution network composed of 8 dilated convolution layers with dilation factors of 1, 2, 4, 8, 16, 32, 64, and 128, respectively, a convolution kernel size of 3, and an output channel number of 64, to obtain the conditional probability distribution parameters of the energy storage demand;

[0063] Based on the conditional probability distribution parameters, a conditional Gaussian distribution is constructed, and Monte Carlo random sampling is performed 100 times within the distribution interval combined with an uncertainty quantification index to obtain multiple samples of the energy storage demand prediction;

[0064] The quantile regression loss function is applied to the multiple samples for calibration calculation to determine the upper and lower bounds of the confidence interval of the prediction distribution, and the energy storage demand prediction distribution is obtained.

[0065] Specifically, the environmental feature vector is processed using the attention mechanism for adaptive weight distribution. The attention mechanism is a technique that dynamically adjusts the weights of each feature based on the importance of the input data. When processing the environmental feature vector, the attention weight value is calculated as the product of a soft-max function and a hyperparameter vector and the environmental feature vector. Specifically, for each environmental feature vector fi (temperature, humidity, wind speed, etc.), its attention weight wi is calculated through the formula wi = softmax( tanh(W·fi+b)) is implemented, where v is the attention vector, W is the weight matrix, and b is the bias term, which are automatically learned through network training. The softmax function converts the original score to a probability distribution, ensuring that the sum of the feature weights is 1. After calculation, the original feature vector is applied with the weight to form a weighted environmental feature vector, so that the model can focus on more important environmental factors.

[0066] The weighted environmental feature vector is processed by residual connection and layer normalization. Residual connection is a technique to prevent the degradation of deep networks by establishing a "shortcut" between the input and output, allowing information to be directly transmitted. In environmental feature processing, residual connection is represented as y = F(x) + x, where x is the input feature, F(x) is the feature converted by the network layer, and y is the output feature. Layer normalization is a standardization process on the same layer of features, which calculates the mean and variance of each layer activation value and adjusts it to prevent the problem of internal covariate shift. The processing process is to first pass the weighted features through the fully connected layer, then add the original features (residual connection), and then perform layer normalization to obtain the feature fusion vector. This vector contains comprehensive information of the environmental features, and the data distribution is more stable, which is convenient for subsequent network processing.

[0067] The feature fusion vector is input into a bidirectional long short-term memory network (BiLSTM). BiLSTM is a special recurrent neural network that can consider both forward and backward information of the sequence. In the design, a 3-layer BiLSTM structure is used, with each hidden layer having a dimension of 256, meaning that the output dimension of each direction of LSTM unit is 128, and after merging, it is 256. BiLSTM controls information flow through the gating mechanism (input gate, forget gate, output gate), which can effectively capture long-distance dependencies. To prevent overfitting, a Dropout layer with a dropout rate of 0.2 is added between each layer of BiLSTM, which randomly shuts down 20% of the neurons to encourage the network to learn more robust feature representations. After processing by the 3-layer BiLSTM, the output is a time series feature representation that fuses forward and backward information, which contains the complete dependency relationship of the environmental features over time.

[0068] TCN is a special one-dimensional convolutional network that can achieve a large receptive field through dilated convolution, which is suitable for processing long sequence data. In the design, 8 dilated convolution layers are used, and the dilated factors are 1, 2, 4, 8, 16, 32, 64 and 128 in turn, forming an exponentially increasing sequence, so that the network can "see" longer historical data. The size of each convolution kernel is 3, which means that each convolution operation considers data at 3 time points. As the network depth increases, the receptive field expands exponentially, eventually covering up to 30 days of historical data. The number of output channels is 64, representing a 64-dimensional feature vector at each time point. These features are converted into conditional probability distribution parameters, including mean μt and variance σt², through the last fully connected layer, which form the basic parameters of the prediction distribution. Building a conditional Gaussian distribution based on the conditional probability distribution parameters is a step to quantify the prediction uncertainty. The conditional Gaussian distribution is a parameterized probability distribution that is completely described by the mean and variance. For each prediction time point t, the model outputs the mean μt and the variance σt², forming the conditional distribution p(yt|xt) = N(μt, σt²), where yt is the energy storage demand prediction value at time t, and xt is the environmental feature input. This distributed prediction has more information than point prediction and can express the uncertainty of the prediction. In order to further consider the influence of environmental parameter uncertainty, the uncertainty quantification index obtained in the previous step is combined to perform Monte Carlo random sampling within the distribution interval. Specifically, based on the parameters of the conditional Gaussian distribution, the variation of the environmental parameter distribution family is considered, and 100 different environmental parameter distributions are randomly sampled. For each distribution, the conditional probability parameters are recalculated to generate 100 energy storage demand prediction samples, which together form the complete prediction distribution considering environmental uncertainty.

[0069] Quantile regression loss function is a special loss function that can predict the conditional distribution of target variable at different quantiles. Unlike mean square error, which only focuses on mean prediction, quantile regression can estimate the complete conditional distribution. During model training, several key quantiles (such as 0.1, 0.5, 0.9) are usually selected for joint optimization. The specific approach is to define the loss function Lτ(y, ) = max(τ(y- ), (1-τ)( -y)) for each quantile τ, where y is the true value, is the predicted value. For multiple samples, quantile regression is performed to calculate the predicted values for each quantile, especially the 0.025 and 0.975 quantiles, which form the 95% confidence interval. Through this calibration process, the model can give a more reliable prediction interval, ultimately forming a complete energy storage demand prediction distribution.

[0070] In a specific embodiment, the process of performing step S104 can specifically include the following steps:

[0071] a target function including energy storage charging and discharging loss cost, energy storage capacity adjustment cost and energy storage response deviation penalty cost is constructed to form a distributionally robust optimization master problem of minimizing the total energy storage operation cost in the worst case, and a mathematical expression of the optimization problem is obtained;

[0072] Based on the energy storage demand prediction distribution, energy storage state of charge constraints, charging and discharging power constraints, charging and discharging switching constraints and distributionally robust chance constraints are set to obtain a constraint condition set of the optimization problem;

[0073] Wasserstein distance metric is introduced to the distributionally robust chance constraint to limit the distance between the probability distribution of the environmental parameter prediction error and the empirical distribution, and a single-layer represented constraint condition is obtained;

[0074] By introducing auxiliary variables and applying the strong duality theorem, equivalent transformation is performed on the target function and the constraint conditions of the optimization problem to obtain a mixed integer quadratic programming model;

[0075] The mixed integer quadratic programming model is applied to a hierarchical decomposition algorithm, the global scheduling problem is decomposed into multiple sub-regional coordination problems, and the alternating direction multiplier method is used to solve each sub-problem to obtain a local optimal solution set;

[0076] The non-active constraint identification technique and the convex relaxation technique are applied to the local optimal solution set, and each sub-regional solution is merged and optimized to obtain the energy storage scheduling instruction sequence for the next 24 hours.

[0077] Specifically, a target function including energy storage charging and discharging loss cost, energy storage capacity adjustment cost and energy storage response deviation penalty cost is constructed to form a distributionally robust optimization master problem of minimizing the total energy storage operation cost in the worst case. The energy storage charging and discharging loss cost represents the energy loss of the energy storage unit in the charging and discharging process, which is usually represented by a quadratic function in the form of μ·P²+ν·P, where P is the charging and discharging power, and μ and ν are loss coefficients determined according to the technical characteristics of the energy storage unit. The energy storage capacity adjustment cost reflects the cost of adjusting the energy storage capacity level, which is described using a piecewise linear function, with different marginal costs corresponding to different capacity intervals. The energy storage response deviation penalty cost is a penalty term for the deviation between the actual response and the demand prediction, which adopts an exponential function λ·exp(ρ·|d|), where d is the deviation value, and λ and ρ are penalty coefficients. The total cost function is obtained by adding the three parts of the cost, and further considering the uncertainty of the environmental parameter distribution, a distributionally robust optimization problem is formed: min{u} max{p∈P} E_{p}[C(u,ξ)], where u is the decision variable (energy storage scheduling instruction), p is the probability distribution of the environmental parameter prediction error, P is the distribution family, ξ is the environmental random variable, and C(u,ξ) is the cost function.

[0078] Setting various constraints based on energy storage demand prediction distribution is an important step in modeling optimization problems. The energy storage state of charge constraint ensures that the state of charge (SOC) of the energy storage unit remains within a safe range, represented as SOC{min}≤SOC(t)≤SOC{max}, where SOC(t) represents the state of charge at time t. This constraint ensures that the energy storage unit is not overcharged or overdischarged. The charge and discharge power constraint limits the charge and discharge power to within the range allowed by the device, represented as -P{max}≤P(t)≤P{max}, where P(t) is the charge and discharge power at time t, with positive values indicating discharge and negative values indicating charge. The charge and discharge switching constraint is used to control the state transition of charge and discharge, introducing a binary variable z(t) to represent the state of charge and discharge at time t (1 for discharge and 0 for charge), and adding the constraint z(t)·P(t)≥0 to ensure that the state and power directions are consistent. The distributionally robust chance constraint handles probabilistic constraints under environmental uncertainty, in the form of Prob{p}{g(u,ξ)≤0}≥1-ε, p∈P, where g(u,ξ) is an environment-dependent constraint function, and ε is the tolerance for violation (usually taken as 0.05). This form of constraint requires that the probability of violating the constraint does not exceed ε under all possible distributions p.

[0079] Introducing the Wasserstein distance metric for the distributionally robust chance constraint is a key step in transforming the problem into a solvable form. The Wasserstein distance is a measure of the difference between two probability distributions, with good geometric interpretation and computational properties. The distributionally robust chance constraint is reformulated as sup_{W(p, )≤δ} Prob{p}{g(u,ξ)≤0}≥1-ε, where is the empirical distribution based on historical data, and δ is the upper bound of the Wasserstein distance, representing the maximum allowed distance between the distribution p and the empirical distribution. This formulation limits the degree of deviation between the predicted error distribution of the environment and the observed historical distribution, making the constraint more robust. Through probability theory and optimization theory, this form of constraint can be transformed into a deterministic constraint, resulting in a single-layer representation of the constraint condition, laying the foundation for subsequent solution.

[0080] By introducing auxiliary variables and applying the strong duality theorem, equivalent transformations of the objective function and constraint conditions of the optimization problem are important technical means for problem solving. The strong duality theorem is an important conclusion in convex optimization theory, which states that when the original problem satisfies certain conditions, the original problem and the dual problem have the same optimal value. For the distributionally robust chance constraint, introduce dual variables λ≥0 and auxiliary variables s, and transform max{p∈P}E{p}[C(u,ξ) into min{λ,s}λδ+E{ [s(ξ)] that satisfies s(ξ) ≥ C(u, ξ) - λd(ξ, ξ̂), where d(ξ, ξ̂) is a distance function in the sample space. Similarly, the distributionally robust chance constraint can also be transformed into a deterministic constraint. After these equivalent transformations, the original two-level optimization problem is transformed into a single-level mixed-integer quadratic programming (MIQP) problem, which contains both continuous variables (e.g., charging / discharging power) and binary variables (e.g., charging / discharging status indicator variables).

[0081] Applying a hierarchical decomposition algorithm to the mixed-integer quadratic programming model is an effective method to handle large-scale optimization problems. Hierarchical decomposition decomposes the global scheduling problem into multiple sub-area coordination problems, each responsible for scheduling a part of the energy storage units. In specific implementation, first define the structure of the master problem and the sub-problems, the master problem coordinates the resource allocation of each sub-area, and the sub-problems optimize the energy storage scheduling within the region given the resource allocation. Solve this decomposition structure through the alternating direction method of multipliers (ADMM), which is an optimization algorithm particularly suitable for distributed optimization problems, and through iterative updates of the original variables, dual variables and Lagrange multipliers, finally converges to the global optimal solution. The update process of ADMM includes three steps: original variable update, dual variable update and multiplier update, and iterates until the convergence condition is met. This decomposition method greatly reduces the computational complexity, making the solution of large-scale energy storage scheduling problems feasible.

[0082] Applying non-active constraint identification technology and convex relaxation technology to the set of local optimal solutions is an important means to improve the solving efficiency. Non-active constraint identification is a preprocessing technique that identifies constraints that will not work at the optimal solution in advance, reducing the problem size. In specific implementation, based on the special structure of the problem and historical solving experience, heuristic rules are designed to identify non-active constraints, such as in periods with large peak-valley price differences, charging / discharging switching constraints usually do not limit. Convex relaxation technology is an effective method to handle non-convex constraints, replacing non-convex constraints (such as binary constraints) with relaxed convex constraints to form a convex optimization problem, and then recovering the relaxed solution to a feasible solution that satisfies the original constraints through heuristic methods. Through these techniques to process the local optimal solutions of each sub-area, then combine them and perform global coordination, finally form a complete energy storage scheduling instruction sequence, including charging / discharging power, energy storage capacity target and other decision instructions for the next 24 hours.

[0083] Taking a certain intelligent microgrid as an example, the microgrid contains multiple distributed energy storage units, and the environmental monitoring system provides environmental data such as temperature and illumination. Based on the environmental monitoring data, the energy storage demand distribution is predicted, including the mean and confidence interval. When building the optimization model, first define the objective function, the energy storage charge and discharge loss cost coefficients are set according to the type of energy storage (such as the loss coefficient μ = 0.02, ν = 0.05 of lithium batteries), the energy storage capacity adjustment cost uses a three-section linear function, the energy storage response deviation penalty parameter λ = 10, ρ = 2. In the constraint condition setting, the state of charge of the energy storage is limited between 20% and 90%, and the upper limit of the charge and discharge power is 80% of the rated power. In the distribution robust opportunity constraint, the tolerance ε is set to 0.05, and the upper limit of the Wasserstein distance δ is determined according to the historical data. By introducing auxiliary variables and applying the strong duality theorem, the distribution robust optimization problem is converted into an MIQP problem, the problem size contains 24 time points, and the decision variables at each time point include charge and discharge power and binary state variables. Due to the large problem size, a hierarchical decomposition algorithm is used to divide the microgrid into 3 sub-regions, each containing part of the energy storage units. The ADMM algorithm is used for iterative solution, and the original and dual residual threshold is set to 10^-4, and the maximum iteration number is 500. For each sub-problem, the non-active constraint identification is applied, which reduces the number of constraints by an average of about 20%. The final energy storage scheduling instruction sequence shows that the energy storage units mainly perform charging operations during periods with sufficient predicted light, and perform discharging operations during periods with insufficient light but high electricity demand, fully utilizing the prediction information provided by the environmental monitoring data and achieving robust scheduling of environmental uncertainty.

[0084] In a specific embodiment, the process of performing step S105 can specifically include the following steps:

[0085] A hierarchical distributed architecture including a coordination layer, an execution layer and a monitoring layer is constructed, and a set of energy storage units is defined, with a directed acyclic graph as the communication topology structure, to obtain a hierarchical distributed control framework;

[0086] The energy storage scheduling instruction sequence is decomposed by a consistent distributed algorithm, and local control instructions are assigned to each energy storage unit according to the principle of minimizing the global coordination cost, to obtain a distributed scheduling instruction set;

[0087] Based on the distributed scheduling instruction set, a state space model including the state of charge and the charge and discharge power of each energy storage unit is established, the prediction time domain is set to 20 time steps, and the control time domain is set to 5 time steps, to obtain the mathematical expression of the model predictive controller;

[0088] A non-periodic sampling data-driven control method is designed for the model predictive controller, the controller gain matrix is obtained by solving a linear matrix inequality, and a control algorithm resistant to communication delay is obtained;

[0089] The voltage, current, temperature, and state of charge are collected in real time by the energy storage unit deployment monitoring module, and the energy efficiency, response time, and control stability are calculated to obtain the operating state data;

[0090] The Lyapunov-Krasovskii functional is constructed based on the operating state data, and the asymptotic stability of the system is verified by proving the negativity of its time derivative. The actual operating data of each energy storage unit are recorded to obtain the energy storage unit operating trajectory data.

[0091] Specifically, a hierarchical distributed architecture including a coordination layer, an execution layer, and a monitoring layer is constructed. This hierarchical architecture is similar to the hierarchical protocol stack in computer networks, with each layer responsible for different functions. The coordination layer is located at the topmost layer and is responsible for the decomposition and issuance of global scheduling strategies, similar to a central control center. The execution layer is located in the middle and receives instructions from the coordination layer and converts them into specific control signals. The monitoring layer is located at the bottom and is directly connected to the energy storage units, collecting operating state data in real time. The set of energy storage units is usually represented as N = {1, 2,..., n}, where n is the total number of energy storage units, each with its own technical characteristics such as capacity, charging and discharging efficiency, etc. The communication topology structure adopts a directed acyclic graph (DAG), which does not contain loops, ensuring one-way information flow and avoiding deadlock and circular dependency problems. In the specific construction process, first, the connection relationship between nodes is determined according to the physical location and communication capability of the energy storage units, then the topological sorting algorithm is applied to ensure the acyclic nature of the graph, and finally the hierarchical distributed control framework is formed. Decomposing the sequence of energy storage scheduling instructions through a consistent convergence distributed algorithm is the key to achieving global coordination. Consistent convergence refers to the process of multiple distributed nodes reaching a consensus decision through local information exchange. In energy storage scheduling, the scheduling instruction u obtained through global optimization needs to be decomposed into local control instructions ui for each energy storage unit, satisfying Σui = u. The decomposition process uses a distributed gradient descent algorithm, and the global coordination cost function is defined as where λi is the weight coefficient, which is adaptively set according to the response capability of the energy storage unit, represents the squared deviation of the local instruction from the average allocation instruction. During the algorithm execution process, each energy storage unit only exchanges information with its neighbor nodes, updates the local control instruction based on the received information and its own local objective function, and gradually converges to a consistent solution through multiple iterations. This distributed algorithm avoids the single-point failure risk of the central controller, improving the robustness and scalability of the system.

[0092] Establishing a state-space model for each energy storage unit based on a distributed scheduling instruction set is fundamental to achieving precise control. A state-space model is a commonly used method for describing dynamic systems, using first-order differential or difference equations to describe the system's internal state and the relationship between external inputs and outputs. For an energy storage unit, the state vector typically includes the state of charge (SOC) and charge / discharge power. The control input is the target charge / discharge power, and the output is the actual SOC. The general form of the state-space model is x(k+1)=Ax(k)+Bu(k), y(k)=Cx(k)+Du(k), where x is the state vector, u is the control input, y is the system output, A, B, C, and D are system matrices, and k represents the discrete time step. A model predictive controller (MPC) is constructed based on this model. The prediction horizon is set to 20 time steps, meaning that the controller predicts the system behavior for the next 20 time steps. The control horizon is set to 5 time steps, meaning that only the control input for the next five time steps is optimized. The mathematical expression of MPC includes an objective function and constraints. The objective function is usually a weighted combination of the sum of the squares of the deviations between the predicted trajectory and the reference trajectory and the sum of the squares of the control input changes.

[0093] In distributed environments, communication delays are a ubiquitous problem that can lead to degraded control performance and even system instability. Aperiodic sampling means that the sampling interval of the control system is not fixed but is dynamically adjusted based on the communication status. Data-driven control methods do not rely on precise system models, but instead learn control strategies directly from historical data. Combining these two techniques, system response data under different communication delay conditions is first collected, and then a delay compensation model is established to predict state changes caused by communication delays. The controller gain matrix K is obtained by solving linear matrix inequalities (LMIs), a special type of convex optimization problem that can effectively handle system stability and performance constraints and is solved using a semidefinite programming algorithm. The resulting control algorithm can adaptively adjust the control strategy to effectively offset the adverse effects of communication delays.

[0094] Deploying a monitoring module in the energy storage unit to collect key parameters in real time is the basis for evaluating the system's operating status. The basic parameters collected by the monitoring module include voltage, current, temperature, and state of charge. The sampling frequency is usually set to 1 second to capture rapidly changing dynamic characteristics. Voltage and current data are used to calculate the instantaneous power and energy flow, temperature data is used to monitor the safety status, and the state of charge reflects the energy storage capacity level. Based on these raw data, multiple performance indicators are calculated: energy efficiency is equal to the output energy divided by the input energy, reflecting the energy conversion loss; response time represents the time delay from receiving the control command to the actual response, usually in milliseconds; control stability is quantified by the standard deviation of the output fluctuation. The smaller the standard deviation, the more stable the control. These calculated operating status data are pre-processed by the edge computing node to filter out noise and outliers, and then uploaded to the coordination layer to form a closed-loop control.

[0095] Constructing Lyapunov-Krasovskii functional based on the operating state data is a theoretical tool for verifying the stability of the system. Lyapunov stability theory is the basis for studying the stability of dynamic systems, while Krasovskii functional is a special Lyapunov function for dealing with delay systems. For the energy storage control system containing communication delay, by verifying this condition, it is ensured that the system can still operate stably in the presence of communication delay. Finally, the state trajectory data of each energy storage unit in the actual operation process is recorded, including the state of charge change, charging and discharging power curve, response time and other data, forming a complete energy storage unit operation trajectory data, which provides practical basis for subsequent strategy evaluation and optimization.

[0096] In a specific embodiment, the process of performing step S106 can specifically include the following steps:

[0097] A multi-dimensional evaluation system including economic indicators, technical indicators and environmental adaptability indicators is constructed, and the performance score of the energy storage unit operation trajectory data is calculated to obtain a feedback evaluation signal;

[0098] Based on the feedback evaluation signal and the environmental characteristics, the energy storage scheduling process is modeled as a Markov decision process, the state space, action space and reward function are defined, and a reinforcement learning optimization framework is obtained;

[0099] A policy network and a value network containing 4 layers of fully connected neural networks are constructed for the reinforcement learning optimization framework, the number of nodes of each layer of the policy network is designed to be 512, 256, 128 and 64, and the activation function is a leaky rectified linear unit, to obtain a deep deterministic policy gradient network structure;

[0100] Based on the energy storage unit operation trajectory data, an experience replay buffer with a capacity of 10000 is established, the batch size is set to 64, the discount factor is set to 0.95, the learning rate is set to 0.0001, and the target network soft update coefficient is set to 0.001, to obtain a training parameter set;

[0101] Ornstein-Uhlenbeck process noise with a mean of 0 and a standard deviation of 0.3 is added to the parameters of the policy network, and the noise standard deviation is decreased according to the training batch, to obtain a parameter space exploration strategy;

[0102] Based on the deep deterministic policy gradient network structure, the training parameter set and the parameter space exploration strategy, the current policy is applied to the actual energy storage system every 100 training batches, the performance data is collected to update the experience replay buffer, and the prediction model hyperparameters, optimization model constraint parameters and controller gain matrix are adjusted through continuous optimization, to obtain an energy storage scheduling strategy.

[0103] Specifically, the continuous improvement of the scheduling strategy is achieved through deep reinforcement learning techniques. First, a multi-dimensional evaluation system containing economic indicators, technical indicators, and environmental adaptability indicators is constructed to calculate the comprehensive performance score of the energy storage unit operation trajectory data. The economic indicators mainly evaluate the cost-effectiveness of energy storage scheduling, including total operating cost, unit storage cost, and marginal scheduling cost, which are quantified by calculating the energy loss, maintenance cost, and price difference during the actual charging and discharging process. The technical indicators evaluate the technical performance of energy storage control, including energy storage response speed, control stability, and scheduling accuracy, among which the scheduling accuracy is calculated by the relative deviation of the actual execution trajectory and the optimization instruction. The environmental adaptability indicators evaluate the adaptability of the energy storage system to environmental changes, including environmental change response, extreme weather adaptability, and environmental prediction deviation robustness, which are calculated by analyzing the sensitivity of the energy storage state to environmental parameter changes. The three types of indicators are combined through weighted average to form the comprehensive performance score, and the weight coefficients are determined by the analytic hierarchy process, which serves as the feedback evaluation signal to guide the subsequent optimization process.

[0104] Modeling the energy storage scheduling process as a Markov Decision Process (MDP) based on the feedback evaluation signal and environmental characteristics is a prerequisite for applying reinforcement learning. Markov Decision Process is a mathematical framework for describing sequential decision-making problems, which includes state space, action space, transition probability, and reward function. In energy storage scheduling, the state space is the joint representation of environmental characteristics and energy storage state, including the current and predicted values of environmental parameters (such as temperature, illumination), and the state of charge, available capacity of energy storage units; the action space is the scheduling strategy parameter space, including the weight coefficients, constraint parameters in the optimization model; the transition probability describes the probability distribution of the system transitioning to the next state given the current state and action, which is implicit in the dynamic process of environmental changes and energy storage responses; the reward function directly uses the comprehensive performance score calculated earlier to measure the immediate return of each state-action pair. Through such modeling, the energy storage scheduling problem is transformed into finding the optimal strategy that maximizes the cumulative reward, forming a theoretical framework for reinforcement learning optimization.

[0105] The construction of a deep deterministic policy gradient network structure is the key to implementing optimization in continuous action space. The deep deterministic policy gradient (DDPG) algorithm is a deep reinforcement learning algorithm suitable for continuous action space, combining the advantages of deep Q-network and policy gradient. DDPG contains two main networks: policy network and value network. The policy network directly outputs deterministic actions, while the value network evaluates the value of state-action pairs. The policy network uses a 4-layer fully connected neural network structure, with node numbers of 512, 256, 128, and 64 from input layer to output layer, respectively. The number of layers gradually decreases, reflecting the process of feature extraction and abstraction. The activation function is the Leaky ReLU, which has a small gradient when the input is negative, avoiding the "dead ReLU" problem and helping to improve training stability. The value network structure is similar to the policy network, but the output layer has a single node that directly outputs the value estimate of the state-action pair. Both networks are trained together through the backpropagation algorithm, and the policy network updates its parameters by maximizing the output of the value network.

[0106] Experience replay is a key technique in deep reinforcement learning, which reduces the correlation between samples and improves data utilization efficiency by storing and reusing historical experience. The buffer capacity is set to 10000, indicating that 10000 transition samples (state, action, reward, next state) can be stored. The oldest sample is removed when a new sample is added, following a first-in, first-out strategy. During training, a batch of 64 samples is randomly selected from the buffer for updating, reducing the correlation between samples. The discount factor is set to 0.95, indicating the discount rate of future rewards. A larger value indicates a greater emphasis on long-term returns. The learning rate is set to 0.0001, controlling the step size of parameter updates. A smaller learning rate helps to improve training stability. The target network soft update coefficient is 0.001, used to slowly update the target network parameters. The target network is a lagging version of the main network, which improves training stability. These parameter settings form a complete set of training parameters, guiding the network training process. Adding noise to the parameters of the policy network is an effective way to implement policy exploration. The Ornstein-Uhlenbeck (OU) process is a commonly used noise model for continuous action space exploration, with the property of mean reversion, i.e., the noise value will gradually return to the mean. The noise generated by the OU process is added to the output action of the policy network, prompting the policy to try different actions and expanding the exploration range. The initial noise standard deviation is set to 0.3, indicating the initial intensity of exploration. As the training batch increases, the noise standard deviation decreases according to a predetermined decay rate, reflecting the gradual transition from exploration to exploitation. This parameter space exploration strategy balances the relationship between exploring new strategies and exploiting known strategies, avoiding local optimal solutions and improving the global optimization ability of reinforcement learning.

[0107] Applying the current policy to the actual energy storage system periodically and collecting performance data is a key step in combining reinforcement learning with actual systems. During training, the current policy is applied to the actual energy storage system every 100 training batches, executed for a period of time (such as a few hours or a day), actual operation data is collected, relevant performance indicators are calculated, and new transition samples are added to the experience replay buffer. This periodic application and data collection forms a closed-loop optimization mechanism for reinforcement learning, constantly adjusting and improving the policy. In the specific optimization process, the parameters of three key components are updated through reinforcement learning: the prediction model hyperparameters (such as network structure, learning rate, etc.), the optimization model constraint parameters (such as tolerance for violations, penalty coefficients, etc.), and the controller gain matrix. The joint optimization of these parameters forms a complete energy storage scheduling strategy, and as reinforcement learning continues, the quality of the strategy improves, and the performance of the system gradually improves.

[0108] The above describes the intelligent environment monitoring data-driven distributed energy storage scheduling prediction method in the embodiments of the present application, and the intelligent environment monitoring data-driven distributed energy storage scheduling prediction system in the embodiments of the present application is described below. Please refer to Figure 2 An embodiment of the intelligent environment monitoring data-driven distributed energy storage scheduling prediction system in the embodiments of the present application includes:

[0109] The processing module 201 is configured to process the temperature, humidity, wind speed, light intensity, and atmospheric pressure data collected by the distributed environment monitoring nodes to obtain an environment data set.

[0110] The decomposition module 202 is configured to decompose the environment data set in multiple levels to obtain an environment feature vector and an uncertainty quantization index.

[0111] The input module 203 is configured to input the environment feature vector and the uncertainty quantization index into a hybrid deep learning model composed of a multi-layer BiLSTM network and a time convolution network to obtain an energy storage demand prediction distribution.

[0112] The conversion module 204 is configured to construct an energy storage scheduling model with a two-layer optimization structure according to the energy storage demand prediction distribution, convert the distribution robust optimization problem into a mixed integer quadratic programming, and obtain an energy storage scheduling instruction sequence.

[0113] The decomposition module 205 is configured to decompose the energy storage scheduling instruction sequence through a hierarchical distributed architecture, distribute global optimization instructions to each energy storage unit, and execute through a non-periodic sampling data-driven model predictive controller to obtain energy storage unit operation trajectory data.

[0114] The optimization module 206 is configured to iteratively optimize the energy storage unit operation trajectory data through a deep deterministic policy gradient algorithm to obtain an energy storage scheduling strategy.

[0115] Through the cooperation of the above-mentioned various components, by processing the temperature, humidity, wind speed, light intensity, and atmospheric pressure data collected by the distributed environment monitoring nodes, a direct mapping relationship between the environmental data and the energy storage demand is established, the data utilization efficiency is improved, and the energy storage scheduling is more in line with the actual environmental conditions. The environmental data set is decomposed at multiple levels to obtain environmental feature vectors and uncertainty quantization indexes, realizing the extraction of deep-level features of environmental data and the accurate quantization of uncertainty. The environmental feature vectors and uncertainty quantization indexes are input into a hybrid deep learning model composed of a multi-layer BiLSTM network and a time convolution network, which fully utilizes the advantages of BiLSTM in capturing long-term dependencies and TCN in processing long sequence data, making the energy storage demand prediction distribution more accurate and reliable. Especially when applying artificial intelligence algorithms, the network structure design of the hybrid deep learning model considers the time sequence characteristics and multi-dimensional features of environmental data, and through the attention mechanism, the feature weights are adaptively allocated, which significantly improves the prediction accuracy. According to the energy storage demand prediction distribution, a two-layer optimization structure of the energy storage scheduling model is constructed, and the distribution robust optimization problem is converted into a mixed integer quadratic programming, effectively dealing with the influence of environmental uncertainty on scheduling and enhancing the robustness of the scheduling strategy. The energy storage scheduling instruction sequence is decomposed through a hierarchical distributed architecture, and the global optimization instruction is distributed to each energy storage unit, which is executed by a model predictive controller driven by non-periodic sampling data, solving the calculation complexity and communication delay problem of traditional centralized control in large-scale distributed systems, and improving the control accuracy and system response speed. Through the deep deterministic policy gradient algorithm, the energy storage unit operation trajectory data is iteratively optimized, realizing the adaptive optimization and continuous improvement of the scheduling strategy, and continuously improving the system performance over time. In the application of artificial intelligence algorithms, the introduction of deep reinforcement learning technology enables the system to learn and improve the strategy from actual operation experience, which is particularly suitable for handling complex problems involving sequential decision-making such as energy storage scheduling. The whole process of intelligent environmental monitoring, demand prediction, scheduling optimization, collaborative control, and strategy optimization is realized.

[0116] The above Figure 2 The intelligent environment monitoring data-driven distributed energy storage scheduling prediction system in the embodiment of the application is described in detail from the perspective of modular functional entities, and the intelligent environment monitoring data-driven distributed energy storage scheduling prediction device in the embodiment of the application is described in detail from the perspective of hardware processing.

[0117] Figure 3is a structural schematic view of an intelligent environment monitoring data driven distributed energy storage scheduling prediction device provided by an embodiment of the present application. The intelligent environment monitoring data driven distributed energy storage scheduling prediction device 300 can have great differences due to different configurations or performances, and can include one or more processors (central processing units, CPUs) 310 (for example, one or more processors) and a memory 320, one or more storage media 330 (for example, one or more mass storage device ends) storing application programs 333 or data 332. The memory 320 and the storage medium 330 can be temporary storage or persistent storage. The programs stored in the storage medium 330 can include one or more modules (not shown in the figure), and each module can include a series of instruction operations in the intelligent environment monitoring data driven distributed energy storage scheduling prediction device 300. Further, the processor 310 can be configured to communicate with the storage medium 330 and execute a series of instruction operations in the storage medium 330 on the intelligent environment monitoring data driven distributed energy storage scheduling prediction device 300 to realize the steps of the intelligent environment monitoring data driven distributed energy storage scheduling prediction method described above.

[0118] The intelligent environment monitoring data driven distributed energy storage scheduling prediction device 300 can also include one or more power supplies 340, one or more wired or wireless network interfaces 350, one or more input and output interfaces 360, and / or one or more operating systems 331, such as Windows Serve, Mac OS X, Unix, Linux, FreeBSD, and the like. Those skilled in the art can understand that the intelligent environment monitoring data driven distributed energy storage scheduling prediction device 300 can also include other components not shown in the figure, such as a bus, an input / output interface, a user interface, a display, a storage medium, and the like. Figure 3 The structure of the intelligent environment monitoring data driven distributed energy storage scheduling prediction device shown in the figure does not constitute a limitation on the intelligent environment monitoring data driven distributed energy storage scheduling prediction device provided by the present application, and can include more or fewer components than shown in the figure, or combine certain components, or different component arrangements.

[0119] The present application also provides a computer readable storage medium, which can be a non-volatile computer readable storage medium or a volatile computer readable storage medium. The computer readable storage medium has instructions stored therein, and when the instructions are run on a computer, the computer executes the steps of the intelligent environment monitoring data driven distributed energy storage scheduling prediction method.

[0120] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described system, system and unit can refer to the corresponding processes in the foregoing method embodiments, which will not be described here.

[0121] The integrated unit, if implemented in the form of a software function unit and sold or used as an independent product, can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application or the entire or part of the technical solutions that essentially contribute to the prior art can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a plurality of instructions for causing an intelligent environment monitoring data-driven distributed energy storage scheduling prediction device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in various embodiments of the present application. The foregoing storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various program code storage media.

[0122] The above embodiments are only used to illustrate the technical solutions of the present application, but not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that the technical solutions recorded in the foregoing embodiments can be modified, or some technical features can be replaced by equivalents; and these modifications or replacements do not make the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A distributed energy storage scheduling prediction method driven by intelligent environmental monitoring data, characterized in that: The method comprises: Process the temperature, humidity, wind speed, light intensity, and atmospheric pressure data collected by distributed environmental monitoring nodes to obtain an environmental data set; The environmental data set is subjected to multi-level decomposition to obtain environmental feature vectors and uncertainty quantification indicators, including: extracting trend components from the environmental data set using a local polynomial fitting method, setting the polynomial order to 3 and using a sliding window technique to obtain trend components of environmental parameters; applying fast Fourier transform and adaptive filtering to the residual data after subtracting the trend component from the environmental data set to obtain periodic components of environmental parameters; performing differential operations on the data after subtracting the trend component and the periodic component from the environmental data set to obtain random fluctuation components of environmental parameters; calculating the slope, curvature and number of inflection points based on the trend component, and calculating the periodic component based on the periodic component. Calculate the main frequency, amplitude, and phase, calculate the fluctuation range, mutation frequency, and duration based on the random fluctuation component, and cascade the three types of feature vectors to obtain the environmental feature vector; establish an empirical distribution function for each environmental parameter in the environmental data set based on historical observation data, and define a confidence interval with a confidence level of 0.95, construct the upper and lower bounds of the distribution function, and obtain a distribution family that includes all possible probability distributions; based on the distribution family, use the Wasserstein distance metric to calculate the upper bound of the distance between the worst-case distribution and the reference distribution, and use the Copula function to capture the correlation structure between multidimensional environmental parameters to obtain the uncertainty quantification index; Inputting the environmental feature vector and uncertainty quantification index into a hybrid deep learning model composed of a multi-layer BiLSTM network and a temporal convolutional network to obtain a predicted distribution of energy storage demand; According to the energy storage demand forecast distribution, a two-layer optimization structure energy storage scheduling model is constructed, and the distributed blue-rod optimization problem is converted into a mixed integer quadratic programming to obtain an energy storage scheduling instruction sequence, including: constructing an objective function including energy storage charging and discharging loss cost, energy storage capacity adjustment cost and energy storage response deviation penalty cost, forming a distributed blue-rod optimization main problem that minimizes the total energy storage operating cost under the worst case, and obtaining a mathematical expression of the optimization problem; setting energy storage charge state constraints, charging and discharging power constraints, charging and discharging switching constraints and distributed blue-rod opportunity constraints based on the energy storage demand forecast distribution to obtain a constraint condition set of the optimization problem; introducing Wasserst for the distributed blue-rod opportunity constraints The ein distance metric is used to limit the distance between the probability distribution of environmental parameter prediction errors and the empirical distribution, thereby obtaining the constraints of a single-layer representation. By introducing auxiliary variables and applying the strong duality theorem, the objective function and constraints of the optimization problem are equivalently transformed to obtain a mixed integer quadratic programming model. A hierarchical decomposition algorithm is applied to the mixed integer quadratic programming model to decompose the global scheduling problem into multiple sub-region coordination problems, and each sub-problem is solved using the alternating direction multiplier method to obtain a set of local optimal solutions. Inactive constraint identification and convex relaxation techniques are applied to the local optimal solution set to merge and optimize the solutions of each sub-region to obtain a sequence of energy storage scheduling instructions for the next 24 hours. The energy storage scheduling instruction sequence is decomposed through a hierarchical distributed architecture, and the global optimization instructions are distributed to each energy storage unit. The instructions are executed through a model predictive controller driven by non-periodic sampling data to obtain the energy storage unit operation trajectory data; The energy storage unit operation trajectory data is iteratively optimized through a deep deterministic policy gradient algorithm to obtain an energy storage scheduling strategy.

2. The distributed energy storage scheduling prediction method driven by intelligent environmental monitoring data according to claim 1 is characterized in that: The temperature, humidity, wind speed, light intensity, and atmospheric pressure data collected by the distributed environmental monitoring nodes are processed to obtain an environmental data set, including: Using a wavelet transform method to remove high-frequency noise from the temperature, humidity, wind speed, light intensity, and atmospheric pressure data collected by the distributed environmental monitoring nodes to obtain denoised environmental data; Normalizing the denoised environmental data to unify the numerical ranges of different types of environmental parameters into the interval [0, 1] to obtain normalized environmental data; Based on time series correlation, an interpolation algorithm is used to repair missing points in the normalized environmental data to obtain complete environmental data; Performing outlier detection and labeling on the complete environmental data by combining the 3σ principle and the density clustering algorithm to obtain labeled environmental data; Reorganizing the marked environmental data into a multi-dimensional time series matrix in chronological order to obtain reconstructed environmental data; The mean, variance, skewness, kurtosis and autocorrelation coefficient of the reconstructed environmental data are calculated, and the statistical features are combined to form the basic components of the environmental feature vector to obtain the environmental data set.

3. The distributed energy storage scheduling prediction method driven by intelligent environmental monitoring data according to claim 1 is characterized in that: The environmental feature vector and uncertainty quantification index are input into a hybrid deep learning model composed of a multi-layer BiLSTM network and a time convolutional network to obtain a predicted distribution of energy storage demand, including: Adopting an attention mechanism to adaptively assign weights to the environmental feature vector, calculating an attention weight value as the product of a soft maximization function, a hyperparameter vector, and the environmental feature vector, to obtain a weighted environmental feature vector; The weighted environment feature vector is processed by residual connection and layer normalization to obtain a feature fusion vector; The feature fusion vector is input into a bidirectional long short-term memory network with a hidden layer dimension of 256 and a layer number of 3, and a dropout layer with a dropout rate of 0.2 is added between each layer to obtain a temporal feature representation that captures the temporal dependency; The temporal feature representation is input into a temporal convolutional network consisting of 8 dilated convolutional layers, with dilation factors of 1, 2, 4, 8, 16, 32, 64, and 128, a convolution kernel size of 3, and 64 output channels, to obtain the conditional probability distribution parameters of energy storage demand; Constructing a conditional Gaussian distribution based on the conditional probability distribution parameters, and performing Monte Carlo random sampling 100 times within the distribution interval in combination with the uncertainty quantification index to obtain multiple groups of samples for energy storage demand prediction; A quantile regression loss function is applied to the multiple groups of samples to perform calibration calculations, determine the upper and lower bounds of the confidence interval of the predicted distribution, and obtain the predicted distribution of energy storage demand.

4. The distributed energy storage scheduling prediction method driven by intelligent environmental monitoring data according to claim 1 is characterized in that: The energy storage scheduling instruction sequence is decomposed through a hierarchical distributed architecture, and the global optimization instructions are distributed to each energy storage unit. The instructions are executed by a model predictive controller driven by non-periodic sampling data to obtain the energy storage unit operation trajectory data, including: A hierarchical distributed architecture consisting of a coordination layer, an execution layer, and a monitoring layer is constructed. A set of energy storage units is defined, and a directed acyclic graph is set as the communication topology to obtain a hierarchical distributed control framework. Decomposing the energy storage scheduling instruction sequence by a distributed algorithm with consistent convergence, allocating local control instructions to each energy storage unit according to the principle of minimizing global coordination cost, and obtaining a distributed scheduling instruction set; Based on the distributed scheduling instruction set, a state space model including the energy storage state of charge and the charge and discharge power is established for each energy storage unit, the prediction time domain is set to 20 time steps, and the control time domain is set to 5 time steps, and a mathematical expression of the model predictive controller is obtained; A non-periodic sampling data driven control method is designed for the model predictive controller, and the controller gain matrix is ​​obtained by solving linear matrix inequalities to obtain a control algorithm that is resistant to communication delays. Deploy a monitoring module in the energy storage unit to collect voltage, current, temperature and state of charge in real time, and calculate energy efficiency, response time and control stability to obtain operating status data; A Lyapunov-Krasovsky functional is constructed based on the operating status data. The asymptotic stability of the system is verified by proving the negative definiteness of its time derivative. The actual operating data of each energy storage unit is recorded to obtain the operating trajectory data of the energy storage unit.

5. The distributed energy storage scheduling prediction method driven by intelligent environmental monitoring data according to claim 1 is characterized in that: The iterative optimization of the energy storage unit operation trajectory data by a deep deterministic policy gradient algorithm to obtain an energy storage scheduling strategy includes: Constructing a multi-dimensional evaluation system including economic indicators, technical indicators and environmental adaptability indicators, performing comprehensive performance scoring calculations on the energy storage unit operation trajectory data, and obtaining feedback evaluation signals; Based on the feedback evaluation signal and environmental characteristics, the energy storage scheduling process is modeled as a Markov decision process, and the state space, action space and reward function are defined to obtain a reinforcement learning optimization framework; A policy network and a value network consisting of a four-layer fully connected neural network are constructed for the reinforcement learning optimization framework. The number of nodes in each layer of the policy network is designed to be 512, 256, 128, and 64, and the activation function is a leaky rectified linear unit, resulting in a deep deterministic policy gradient network structure. Based on the energy storage unit operation trajectory data, an experience replay buffer with a capacity of 10,000 is established, and the batch size, discount factor, learning rate, and target network soft update coefficient are set to 64, 0.95, 0.0001, and 0.001, respectively, to obtain a training parameter set. Adding Ornstein-Uhlenbeck process noise with a mean of 0 and a standard deviation of 0.3 to the parameters of the policy network, and decreasing the standard deviation of the noise according to the training batch to obtain a parameter space exploration strategy; Based on the deep deterministic policy gradient network structure, the training parameter set, and the parameter space exploration strategy, the current strategy is applied to the actual energy storage system every 100 training batches, performance data is collected to update the experience replay buffer, and the energy storage scheduling strategy is obtained by continuously optimizing and adjusting the prediction model hyperparameters, optimizing the model constraint parameters, and optimizing the controller gain matrix.

6. A distributed energy storage scheduling prediction system driven by intelligent environmental monitoring data, characterized in that: The method for implementing a distributed energy storage scheduling prediction method driven by intelligent environmental monitoring data according to any one of claims 1 to 5, wherein the distributed energy storage scheduling prediction system driven by intelligent environmental monitoring data comprises: The processing module is used to process the temperature, humidity, wind speed, light intensity, and atmospheric pressure data collected by the distributed environmental monitoring nodes to obtain an environmental data set; The decomposition module is used to perform multi-level decomposition on the environmental data set to obtain environmental feature vectors and uncertainty quantitative indicators, including: extracting trend components from the environmental data set using a local polynomial fitting method, setting the polynomial order to 3 and using a sliding window technique to obtain trend components of environmental parameters; applying fast Fourier transform to the residual data after subtracting the trend component from the environmental data set and combining it with adaptive filtering to obtain periodic components of environmental parameters; performing differential operation on the data after subtracting the trend component and the periodic component from the environmental data set to obtain random fluctuation components of environmental parameters; calculating the slope, curvature and number of inflection points based on the trend component, and calculating the periodic components based on the trend component. The main frequency, amplitude, and phase of the periodic component are calculated, and the fluctuation range, mutation frequency, and duration are calculated based on the random fluctuation component. The three types of characteristic vectors are cascaded to obtain the environmental characteristic vector; for each environmental parameter in the environmental data set, an empirical distribution function is established based on historical observation data, and a confidence interval with a confidence level of 0.95 is defined. The upper and lower bounds of the distribution function are constructed to obtain a distribution family containing all possible probability distributions; based on the distribution family, the Wasserstein distance metric is used to calculate the upper bound of the distance between the worst-case distribution and the reference distribution, and the Copula function is used to capture the correlation structure between multidimensional environmental parameters to obtain the uncertainty quantification index; An input module, configured to input the environmental feature vector and uncertainty quantification index into a hybrid deep learning model composed of a multi-layer BiLSTM network and a temporal convolutional network to obtain a predicted distribution of energy storage demand; A conversion module is used to construct a two-layer optimization structure of an energy storage scheduling model according to the energy storage demand forecast distribution, convert the distributed blue-rod optimization problem into a mixed integer quadratic programming, and obtain an energy storage scheduling instruction sequence, including: constructing an objective function including energy storage charging and discharging loss cost, energy storage capacity adjustment cost and energy storage response deviation penalty cost, forming a distributed blue-rod optimization main problem that minimizes the total energy storage operating cost under the worst case, and obtaining a mathematical expression of the optimization problem; setting energy storage charge state constraints, charging and discharging power constraints, charging and discharging switching constraints and distributed blue-rod opportunity constraints based on the energy storage demand forecast distribution, and obtaining a constraint condition set of the optimization problem; introducing Wasse into the distributed blue-rod opportunity constraint The rstein distance metric is used to limit the distance between the probability distribution of environmental parameter prediction errors and the empirical distribution, thereby obtaining the constraints of a single-layer representation. By introducing auxiliary variables and applying the strong duality theorem, the objective function and constraints of the optimization problem are equivalently transformed to obtain a mixed integer quadratic programming model. A hierarchical decomposition algorithm is applied to the mixed integer quadratic programming model to decompose the global scheduling problem into multiple sub-region coordination problems, and each sub-problem is solved using the alternating direction multiplier method to obtain a set of local optimal solutions. Inactive constraint identification and convex relaxation techniques are applied to the local optimal solution set to merge and optimize the solutions of each sub-region to obtain a sequence of energy storage scheduling instructions for the next 24 hours. a decomposition module for decomposing the energy storage scheduling instruction sequence through a hierarchical distributed architecture, allocating global optimization instructions to each energy storage unit, and executing them through a model predictive controller driven by non-periodic sampling data to obtain energy storage unit operation trajectory data; The optimization module is used to iteratively optimize the operation trajectory data of the energy storage unit through a deep deterministic policy gradient algorithm to obtain an energy storage scheduling strategy.

7. A distributed energy storage scheduling prediction device driven by intelligent environmental monitoring data, characterized in that: The method comprises a memory and a processor, wherein the memory stores a computer program that can be run on the processor, and when the processor executes the computer program, the distributed energy storage scheduling prediction method driven by intelligent environmental monitoring data according to any one of claims 1 to 5 is implemented.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the processor is enabled to execute the distributed energy storage scheduling prediction method driven by intelligent environmental monitoring data according to any one of claims 1 to 5.

Images