Optical storage direct flexible system short-term load prediction method based on multilevel data association fusion

Abstract

The invention discloses a short-term load prediction method for an optical storage direct flexible system based on multilevel data association fusion, and the method comprises the steps: collecting and processing the uncertain load data, such as photovoltaic output power and building electrical load, through constructing a multi-source heterogeneous data collection and preprocessing system. Dynamic weighted fusion and correlation analysis are adopted to calculate dynamic correlation weights among multi-source data, multi-scale feature extraction is carried out, and a fusion data set is constructed. And developing a multi-modal attention fusion prediction model, designing a space-time joint attention mechanism and the like, and realizing explicit modeling of long-range correlation and nonlinear interaction characteristics in high-dimensional data. And deploying an intelligent load prediction and decision support platform, integrating a prediction model to an energy management system, generating a comprehensive report in combination with the dynamic electricity price and an energy storage control strategy, and interacting with an SCADA system in real time and issuing the strategy through an OPC UA protocol. According to the invention, the accuracy of load prediction and the stability of system operation are effectively improved, and optimal management and efficient utilization of energy are realized.

Description

Technical Field

[0001] The present invention relates to the field of energy management, and particularly to a short-term load forecasting method for a photovoltaic-storage-direct-current-flexible (PV-SD-F) system with multi-level data association and fusion. Background Art

[0002] In the context of the global energy transition, the PV-SD-F system, as an innovative energy management solution, has been gradually attracting wide attention. However, there are still many deficiencies in the existing technologies for short-term load forecasting in PV-SD-F systems.

[0003] Traditional energy management systems have limitations in data collection in the field of PV-SD-F. They mainly focus on collecting single-type energy data, such as only paying attention to photovoltaic output power or building electricity load, and lack data collection for emerging energy forms such as electric vehicle charging load. This leads to the lack of comprehensiveness and diversity of data and cannot accurately reflect the real situation of energy flow in the PV-SD-F system. In the data processing link, for the multi-source heterogeneous data in the PV-SD-F system, traditional methods lack efficient time series alignment, outlier cleaning, and normalization processing strategies. As a result, the data quality is poor and it is difficult to meet the requirements of subsequent accurate analysis and modeling. For load forecasting, existing methods perform poorly in PV-SD-F systems. They cannot fully explore the deep associations and dynamic change laws among the data of various parts such as photovoltaic, energy storage, direct current, and flexible load, and it is difficult to accurately predict short-term load changes, resulting in unreasonable energy allocation and unstable system operation. In terms of decision support, traditional systems are difficult to formulate precise and flexible electricity price strategies and energy storage control strategies based on the real-time state of the PV-SD-F system and the complex energy market dynamics. They cannot generate comprehensive and highly targeted comprehensive reports and cannot provide strong support for the optimal operation of the PV-SD-F system.

[0004] In summary, the existing technical means have obvious deficiencies in data collection, processing, load forecasting, and decision support of PV-SD-F systems, and it is difficult to meet the requirements of the efficient and stable operation of PV-SD-F systems. Summary of the Invention

[0005] To solve the above problems, the present invention proposes a short-term load forecasting method for a PV-SD-F system with multi-level data association and fusion. The specific steps are as follows, and it is characterized in that:

[0006] Step 1: Construct a multi-source heterogeneous data collection and preprocessing system. Deploy a distributed sensor network in the PV-SD-F system to collect uncertainty load data such as photovoltaic output power, building electricity load, and electric vehicle charging load in real time; perform time series alignment, outlier cleaning, and normalization processing on the original data through edge computing nodes, and establish a multi-dimensional load data set containing spatio-temporal features;

[0007] Step 2: Implement dynamic weighted fusion and correlation analysis, and use an improved grey correlation algorithm to calculate the dynamic correlation weights among multi-source data such as photovoltaic, building load, and electric vehicle charging; combine an adaptive mode decomposition strategy (such as VMD-CEEMDAN hybrid decomposition) to extract multi-scale features from non-stationary load sequences, and construct a fusion dataset with hierarchical feature representation;

[0008] Step 3: Develop a multi-modal attention fusion prediction model, design a spatio-temporal joint attention mechanism based on the Transformer architecture, and optimize the computational efficiency of the self-attention module through a hierarchical distillation strategy; introduce a bidirectional GRU network in the encoder layer to capture temporal dependencies, and embed a dynamic coupling analysis module in the decoder layer to achieve explicit modeling of long-range correlations and non-linear interaction features in high-dimensional data;

[0009] Step 4: Deploy an intelligent load prediction and decision support platform, integrate the trained prediction model into the energy management system (EMS); combine dynamic electricity price strategies and energy storage control strategies to generate a comprehensive report including power fluctuation warnings, energy storage scheduling suggestions, and demand response plans, and realize real-time data interaction and strategy issuance with the SCADA system through the OPC UA protocol.

[0010] The short-term load prediction method for the photovoltaic-storage-direct-current-flexible system with multi-level data association and fusion of the present invention, beneficial effects, the technical effects of the present invention are as follows:

[0011] 1. The present invention constructs a multi-source heterogeneous data acquisition and preprocessing system, which can collect various uncertain load data in the photovoltaic-storage-direct-current-flexible system in real time and accurately, and perform effective preprocessing, providing a high-quality data basis for subsequent analysis and prediction.

[0012] 2. The present invention implements dynamic weighted fusion and correlation analysis, which can calculate the dynamic correlation weights among multi-source data more precisely, extract multi-scale features, and construct a fusion dataset with hierarchical feature representation, thereby improving the accuracy and adaptability of prediction.

[0013] 3. The multi-modal attention fusion prediction model developed by the present invention, through innovative designs such as a spatio-temporal joint attention mechanism, a hierarchical distillation strategy, a bidirectional GRU network, and a dynamic coupling analysis module, can explicitly model long-range correlations and non-linear interaction features in high-dimensional data, significantly improving the prediction accuracy of the short-term electricity load of the photovoltaic-storage-direct-current-flexible system. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 is a flow chart of the present invention;

[0015] Figure 2 is a schematic diagram of the construction of the multi-source heterogeneous data acquisition and preprocessing system of the present invention;

[0016] Figure 3 This is a schematic diagram of the deployment of the intelligent load forecasting and decision-making support platform for the present invention. Detailed implementation manners

[0017] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific implementation manners:

[0018] The present invention proposes a short-term load forecasting method for a photovoltaic-storage-direct-current-soft system with multi-level data association and fusion. The flowchart of the invention is as Figure 1 shown. The steps of the present invention will be introduced in detail below.

[0019] Step 1: Construct a multi-source heterogeneous data acquisition and preprocessing system. Deploy a distributed sensor network in the photovoltaic-storage-direct-current-soft system to collect uncertainty load data such as photovoltaic output power, building electricity load, and electric vehicle charging load in real time; perform time series alignment, outlier cleaning, and normalization processing on the original data through edge computing nodes, establish a multi-dimensional load data set containing spatio-temporal features, and construct a schematic diagram of the multi-source heterogeneous data acquisition and preprocessing system as Figure 2 shown;

[0020] Step 1.1, deploy distributed sensors, including photovoltaic system sensors, building electricity load sensors, electric vehicle charging load sensors, and energy storage system monitoring sensors.

[0021] In the deployment of photovoltaic system sensors, the light intensity sensor measures the solar radiation level, the temperature sensor monitors the surface temperature of the photovoltaic panel, the current and voltage sensors monitor the DC output power of the photovoltaic module, and the power meter measures the AC output power of the photovoltaic inverter.

[0022] In the deployment of building electricity load sensors, the smart meter records the power consumption at different time periods, the current and voltage transformers are used to monitor the power flow direction of the distribution circuit, and the environmental sensors include: temperature and humidity, lighting intensity, CO 2 concentration, etc., for evaluating the building load characteristics.

[0023] In the deployment of electric vehicle charging load sensors, the current and voltage sensors of the charging pile monitor and collect the load change characteristics, charging power, and time of different charging modes in real time, the vehicle identification module obtains the vehicle type and the battery SOC state, and the charging station management system.

[0024] In the deployment of energy storage system monitoring sensors, the battery management system collects the battery SOC, SOH, temperature, current, and voltage, and the energy management system records the energy storage charging and discharging strategies and the dynamic interaction with the load.

[0025] At the same time, a wireless sensor network is adopted for data transmission, and data aggregation, preprocessing are carried out through an edge computing gateway, and the data is transmitted to the upper-layer system.

[0026] Step 1.2, Time series alignment. The data alignment steps include:

[0027] Step 1.2.1, Timestamp standardization, unifying the timestamp formats of all data sources;

[0028] Step 1.2.2, Select the target time step, set a unified time step to ensure that all data points fall within the same time interval.

[0029] Step 1.2.3, Interpolation to fill in missing data. For irregular missing values, an interpolation algorithm is used for filling.

[0030] Step 1.2.4, Downsampling based on window averaging. For oversampled data, the sliding window averaging method is adopted.

[0031] Step 1.3, Outlier detection and cleaning. First, a reasonable data range is set, then the Z-score standardization statistical method is used to identify outliers, and the outliers are removed. Finally, the data is subjected to sliding filtering and normalization to establish a multi-dimensional load dataset containing spatio-temporal characteristics.

[0032] Step 2: Implement dynamic weighted fusion and correlation analysis. The improved grey correlation degree algorithm is used to calculate the dynamic correlation weights among multi-source data such as photovoltaic, building load, and electric vehicle charging; combined with the VMD-CEEMDAN hybrid decomposition strategy, multi-scale feature extraction is carried out on non-stationary load sequences to construct a fusion dataset with hierarchical feature expression;

[0033] Step 2.1, Use the improved grey correlation degree algorithm to calculate the dynamic correlation weights among multi-source data such as photovoltaic, building load, and electric vehicle charging. Calculate the dynamic correlation weights between different loads under a dynamic time window to provide a basis for subsequent feature extraction and data fusion.

[0034] Step 2.1.1, Let the reference sequence of the grey correlation degree be X 0 ={x 0 (1),x 0 (2),...,x 0 (n)}, X 0 represents the short-term electricity load data set of the photovoltaic-storage-direct-current flexible system, and x 0 (n) represents the short-term electricity load of the photovoltaic-storage-direct-current flexible system at time series n. The comparison sequence is X i ={x i (1),x i (2),...,x i (n)}, i = 1, 2,..., m, xi It represents the data such as photovoltaic power and electric vehicle load collected by the sensors in Step 1. m is the dimension of the data collected by all sensors in Step 1. Then the grey correlation coefficient is defined as follows:

[0035]

[0036] Among them, ξ i (k) is the grey correlation coefficient at time point k. ρ is the resolution coefficient, taking 0.5 to balance the data distribution. |x 0 (k) - x i (k)| represents the absolute difference at data point k. represents the minimum absolute difference in all data sequences, represents the maximum absolute difference in all data sequences. The grey correlation degree of the whole sequence can be obtained through weighted average:

[0037]

[0038] Among them, Γ i represents the global correlation degree between the comparison sequence X i and the reference sequence X 0 , with a value range of [0, 1]. The closer it is to 1, the stronger the correlation.

[0039] Step 2.1.2 Dynamic sliding window calculation

[0040] The data is segmented and calculated using a sliding time window, and it is continuously updated as time goes by. Let the length of the time series be N, and the length of the sliding window ω is selected. Within each window, the local grey correlation degree is calculated:

[0041]

[0042] Among them, Γ i t represents the local grey correlation degree at the t-th moment, which can reflect the short-term dynamic characteristics.

[0043] Step 2.1.3 Calculate the dynamic correlation weight

[0044] Based on the improved grey correlation degree Γ i t , calculate the dynamic correlation weight of different data sources at each time point:

[0045]

[0046] Among them, W i tDenote the dynamic weights of data sources such as photovoltaic, building load, and electric vehicle charging load at time t. This weight is used for subsequent multi-source data fusion to improve the adaptive ability of the prediction model. Denote the normalization factor to ensure that the sum of the dynamic weights of all data sources is 1.

[0047] Step 2.2 uses VMD-CEEMDAN for multi-scale feature extraction.

[0048] Step 2.2.1, adopt a weighted data fusion strategy to form the optimized input data X * (t):

[0049]

[0050] Take X * (t) as the input for VMD-CEEMDAN decomposition.

[0051] Step 2.2.2, decompose the input signal X * (t) into K narrow-band sub-sequence modes:

[0052]

[0053] VMD ensures that the frequency bands of each mode do not overlap by solving the following variational optimization problem:

[0054]

[0055] where u k (t) is the K-th mode component. ω k is the central frequency of the K-th mode component. H(u k ) is the Hilbert transform of the K-th mode component to obtain the analytic representation of the signal. is the second norm, representing the least square error of the optimization objective function, and j is the imaginary unit.

[0056] VMD is solved by the alternating direction method of multipliers to obtain K mode components u k (t), and each component contains features within a specific frequency range.

[0057] Step 2.2.3 Complete Ensemble Empirical Mode Decomposition CEEMDAN

[0058] After VMD decomposition, each mode component u k (t) still contains fine-grained non-stationary features. To further extract key patterns, perform CEEMDAN decomposition on each u k (t):

[0059]

[0060] Among them, IMF k,j (t) is the intrinsic mode component obtained by CEEMDAN decomposition, J is the decomposition order, and r k (t) is the residual term. By adding adaptive noise, CEEMDAN avoids the mode mixing problem in the traditional EMD method and makes the decomposition result more stable.

[0061] Step 2.3. Combine dynamic weights to optimize feature fusion

[0062] Finally, obtain the subsequence set after multi-scale decomposition:

[0063] D = {IMF k,j (t)|k = 1, 2,..., K; j = 1, 2,..., J}

[0064] Different IMFs k,j (t) have different degrees of influence on the target variable. In order to optimize data fusion, the dynamic grey relational degree is used again to calculate the weights of each subsequence and construct the final fusion feature dataset:

[0065]

[0066] Among them, represents the final fusion dataset combined with dynamic weights, and W t represents the normalized dynamic weight, indicating the relative influence of the j-th IMF of the k-th VMD component on the target variable.

[0067] Step 3: Develop a multi-modal attention fusion prediction model, design a spatio-temporal joint attention mechanism based on the Transformer architecture, and optimize the computational efficiency of the self-attention module through a hierarchical distillation strategy; introduce a bidirectional GRU network in the encoder layer to capture temporal dependencies, and embed a dynamic coupling analysis module in the decoder layer to achieve explicit modeling of long-range correlations and non-linear interaction features in high-dimensional data;

[0068] Step 3.1 Design a spatio-temporal joint attention mechanism

[0069] Adopt a spatio-temporal joint attention mechanism for feature fusion. This mechanism is based on the architecture of the Transformer model and introduces a spatio-temporal joint attention module to capture long-term dependencies in different modalities and time series respectively.

[0070] The spatio-temporal joint attention mechanism performs weighted summation on the multi-modal input data at each time step through the self-attention mechanism, thereby effectively focusing on the time steps and modalities that are most important for prediction, ensuring that when the model processes different types of loads, it can weight different input signals according to the actual situation and obtain accurate short-term electricity load predictions.

[0071] Step 3.2 Optimize the computational efficiency of the self-attention module using a hierarchical distillation strategy

[0072] The distillation strategy makes the feature representations at higher levels more concise and efficient by gradually reducing the complexity of attention calculations. By gradually learning simplified attention patterns, unnecessary computations of the model are reduced, thereby improving the overall computational efficiency.

[0073] Through hierarchical distillation, unnecessary computational amounts in the self-attention mechanism are reduced, the time costs of model training and inference are lowered, enabling the system to respond in real time to changing load data.

[0074] Step 3.3 Introduce a bidirectional GRU network in the encoder layer

[0075] A bidirectional GRU network is introduced in the encoder layer of the Transformer. The bidirectional GRU can consider temporal information in both the forward and backward directions simultaneously, enhancing the model's sensitivity to historical load data and future load changes. The bidirectional GRU network can obtain the temporal features of load data from both the forward and backward directions simultaneously. In this way, the dynamic evolution pattern of load data can be better understood, thereby improving the prediction accuracy of load changes.

[0076] Step 3.4 Embed a dynamic coupling analysis module in the decoder layer

[0077] In the decoder layer, a dynamic coupling analysis module is added to capture the non-linear interaction relationships between different data sources. This module can dynamically capture the non-linear interaction patterns between loads, combine the temporal information in the forward and backward directions, explicitly model the interaction features between load data, improve the expressiveness of the model, and make the final prediction results more accurate.

[0078] Step 3.5 Explicitly model long-range correlations

[0079] With the support of the spatio-temporal joint attention mechanism, the model can explicitly model long-range correlations in high-dimensional data. Through the Transformer architecture, especially by introducing the self-attention mechanism, the model can pay attention to long-range dependence features across time steps through the global attention mechanism, enabling the mutual influences between multiple data sources such as photovoltaic, electric vehicle charging loads, and building loads to be fully learned.

[0080] Step 4: Deploy an intelligent load prediction and decision support platform, integrate the trained prediction model into the Energy Management System (EMS); combine dynamic electricity price strategies and energy storage control strategies to generate a comprehensive report including power fluctuation warnings, energy storage scheduling suggestions, and demand response plans, and realize real-time data interaction and strategy distribution with the SCADA system through the OPC UA protocol. The schematic diagram of the deployed intelligent load prediction and decision support platform is as Figure 3 shown.

[0081] Step 4.1 Integrate the prediction model into the energy management system

[0082] Integrate the trained and optimized load prediction model into the energy management system to ensure that the model can perform real-time prediction and dynamic decision-making in the actual operating environment. The EMS is responsible for controlling and dispatching the energy resources in the system, including photovoltaic, battery energy storage, electric vehicle charging stations, etc. By introducing the load prediction model into the EMS, the system can perform real-time load demand prediction and decision support. The system predicts future load demands based on multi-source data such as photovoltaic power generation, electric vehicle charging load, and building load, and makes short-term predictions about their changes. The EMS combines the prediction results, evaluates different operation strategies, and formulates control strategies such as energy storage dispatching, demand response, and load balancing.

[0083] Step 4.2 Combine the dynamic electricity price strategy with the energy storage control strategy. The system introduces the dynamic electricity price strategy and the energy storage control strategy.

[0084] Dynamic electricity price strategy: According to the real-time electricity price fluctuations of the power grid, the system can adjust the load demand.

[0085] Energy storage control strategy: Combining the energy storage capacity and battery charge and discharge requirements in the photovoltaic-storage-direct-current flexible system, the system adjusts the charge and discharge strategies of the energy storage device based on the load prediction to balance the load demand of the power grid. The energy storage system can not only regulate the energy supply but also cope with the uncertainty and volatility of photovoltaic power generation to ensure the stable operation of the system.

[0086] Step 4.3 Generate a comprehensive report

[0087] Through the prediction and decision-making of real-time load data, the system can generate a comprehensive report containing the following content:

[0088] Power fluctuation warning: Based on the load prediction and the power grid status, the system warns in advance of possible power fluctuation situations and proposes countermeasures. By predicting the load demand fluctuations, the system takes corresponding dispatching measures when the power grid load is too high or too low to reduce the power grid burden or balance the system.

[0089] Energy storage dispatching suggestions: Based on the status of the energy storage system and the load prediction, the system gives energy storage dispatching suggestions, such as the recommended charging or discharging time and charging power. By optimizing the energy storage dispatching strategy, ensure that the battery operates in an efficient and safe state.

[0090] Demand response plan: For different peak power demand periods, the system automatically generates a demand response plan to adjust building loads, charging loads, etc. to balance the load demand of the power grid and achieve load management and power savings.

[0091] Step 4.4 realizes real-time data interaction with the SCADA system and policy distribution

[0092] To ensure the real-time nature of load forecasting and decision-making strategies, the system conducts data interaction with the SCADA system through the OPC UA protocol. The OPC UA protocol enables the EMS to achieve efficient data transmission and interaction with the SCADA system, obtain data from the sensor network in real time, and feedback the prediction results and scheduling strategies to the SCADA system. Through the SCADA system, the EMS distributes adjustment instructions in real time and executes control measures such as energy storage charging and discharging and demand response to ensure that the load and power supply status of the power grid and the PV-BESS-DR system always remain within a reasonable range.

[0093] The above are only the preferred embodiments of the present invention, and do not impose any other form of limitation on the present invention. Any modification or equivalent change made based on the technical essence of the present invention still falls within the scope of protection required by the present invention.

Claims

1. A short-term load forecasting method for a photovoltaic-storage-direct-flexible system based on multi-level data association fusion, the specific steps are as follows, and the characteristics are as follows: Step 1: Build a multi-source heterogeneous data collection and preprocessing system, deploy a distributed sensor network in the solar-storage-direct-flexible system, and collect real-time uncertain load data such as photovoltaic output power, building power load, and electric vehicle charging load; use edge computing nodes to align the time series, clean up outliers, and normalize the raw data to establish a multi-dimensional load data set with spatiotemporal characteristics; Step 2: Implement dynamic weighted fusion and correlation analysis, and use the improved grey correlation algorithm to calculate the dynamic correlation weights between multi-source data such as photovoltaic, building load, and electric vehicle charging; Combined with VMD-CEEMDAN hybrid decomposition strategy, multi-scale feature extraction is performed on non-stationary load series to construct a fusion dataset with hierarchical feature expression; Step 3: Develop a multimodal attention fusion prediction model, design a spatiotemporal joint attention mechanism based on the Transformer architecture, and optimize the computational efficiency of the self-attention module through a hierarchical distillation strategy; introduce a bidirectional GRU network at the encoder layer to capture temporal dependencies, and embed a dynamic coupling analysis module at the decoder layer to achieve explicit modeling of long-range correlations and nonlinear interaction features in high-dimensional data; Step 4: Deploy an intelligent load forecasting and decision support platform, integrate the trained forecasting model into the energy management system (EMS); combine the dynamic electricity price strategy and energy storage control strategy to generate a comprehensive report including power fluctuation warning, energy storage scheduling suggestions, and demand response solutions, and realize real-time data interaction and policy delivery with the SCADA system through the OPC UA protocol.

2. The short-term load forecasting method for a PV-storage-direct-flexible system based on multi-level data association fusion according to claim 1 is characterized by: The construction of a multi-source heterogeneous data acquisition and preprocessing system in step 1 can be expressed as: Step 1.1, deploy distributed sensors, including photovoltaic system sensors, building power load sensors, electric vehicle charging load sensors, and energy storage system monitoring sensors; In the deployment of sensors for photovoltaic systems, light intensity sensors measure the solar radiation level, temperature sensors monitor the surface temperature of photovoltaic panels, current and voltage sensors monitor the DC output power of photovoltaic modules, and power meters measure the AC output power of photovoltaic inverters; In the deployment of building electricity load sensors, smart meters record electricity consumption in different time periods, current and voltage transformers are used to monitor the power flow of the distribution circuit, and environmental sensors include: temperature and humidity, lighting intensity, CO2 concentration, etc., which are used to evaluate building load characteristics; In the deployment of electric vehicle charging load sensors, the current and voltage sensors of the charging piles monitor and collect the load change characteristics of different charging modes in real time, charging power and time, and the vehicle identification module obtains the vehicle type, battery SOC status, and charging station management system; In the deployment of energy storage system monitoring sensors, the battery management system collects battery SOC, SOH, temperature, current, and voltage, and the energy management system records the energy storage charging and discharging strategy and the dynamic interaction with the load; At the same time, wireless sensor networks are used for data transmission, and edge computing gateways are used to aggregate and pre-process data and transmit it to the upper system; Step 1.2, time series alignment, the data alignment steps include: Step 1.2.1: Standardize timestamps and unify the timestamp formats of all data sources; Step 1.2.2, select the target time step, set a uniform time step, and ensure that all data points fall in the same time interval; Step 1.2.3, interpolation to complete missing data. For irregular missing values, interpolation algorithms are used to fill them. Step 1.2.4, based on window average downsampling, for oversampled data, the sliding window average method is used; Step 1.3, outlier detection and cleaning, first set a reasonable data range, then use the Z-score standardized statistical method to identify outliers, remove outliers, and finally perform sliding filtering and normalization on the data to establish a multidimensional load data set containing spatiotemporal characteristics.

3. The short-term load forecasting method for a PV-storage-direct-flexible system based on multi-level data association fusion according to claim 1 is characterized by: The implementation of dynamic weighted fusion and correlation analysis in step 2 can be expressed as follows: Step 2.1, using the improved grey correlation algorithm to calculate the dynamic correlation weights between multi-source data such as photovoltaic, building load, and electric vehicle charging, and calculating the dynamic correlation weights between different loads under the dynamic time window to provide a basis for subsequent feature extraction and data fusion; Step 2.1.1, let the reference sequence of grey correlation be X0 = {x0(1), x0(2), ..., x0(n)}, X0 represents the short-term power load data set of the PV-storage-DC-flexible system, x0(n) represents the short-term power load of the PV-storage-DC-flexible system at time n, and the comparison sequence is X i ={x i (1),x i (2),...,x i (n)}, i = 1, 2, ..., m, x i represents the photovoltaic power, electric vehicle load and other data collected by the sensor in step 1, m is the dimension of the data collected by all sensors in step 1, and the grey correlation coefficient is defined as follows: Among them, ξ i (k) is the grey relational coefficient at time point k; ρ is the resolution coefficient, which is taken as 0.5 to balance the data distribution; |x0(k)-x i (k)| represents the absolute difference at data point k; represents the minimum absolute difference among all data sequences, Represents the maximum absolute difference in all data sequences. The grey relational degree of the entire sequence can be obtained by weighted average: Among them, Γ i Represents the comparison sequence X i The global correlation with the reference sequence X0 has a value range of [0,1]. The closer it is to 1, the stronger the correlation. Step 2.1.2 Dynamic sliding window calculation The data is segmented and calculated using a sliding time window, which is continuously updated over time. Let the length of the time series be N, and the length of the sliding window be ω. In each window, the local grey relational degree is calculated: in, It represents the local grey relational degree at the tth moment, which can reflect the short-term dynamic characteristics; Step 2.1.3 Calculate dynamic association weight Based on improved grey relational Calculate the dynamic correlation weights of different data sources at each time point: Among them, W i t It represents the dynamic weight of data sources such as photovoltaic, building load, and electric vehicle charging load at time t. This weight is used for subsequent multi-source data fusion to improve the adaptive ability of the prediction model. Represents the normalization factor, ensuring that the sum of the dynamic weights of all data sources is 1; Step 2.2 Use VMD-CEEMDAN for multi-scale feature extraction Step 2.2.1: Use weighted data fusion strategy to form optimized input data X * (t): X * (t) as the input of VMD-CEEMDAN decomposition; Step 2.2.2, input signal X * (t) is decomposed into K narrowband subsequence modes: VMD ensures that the frequency bands of each mode do not overlap by solving the following variational optimization problem: Among them, u k (t) is the Kth modal component; ω k is the center frequency of the Kth modal component; H(u k ) is the Hilbert transform of the Kth modal component, and the analytical representation of the signal is obtained; is the two-norm, representing the least squares error of the optimization objective function, and j is a complex unit; VMD is solved by the alternating direction multiplier method to obtain K modal components u k (t), each component contains features within a specific frequency range; Step 2.2.3 Fully integrated empirical mode decomposition CEEMDAN After VMD decomposition, each modal component u k (t) still contains fine-grained non-stationary features; in order to further extract key patterns, for each u k (t) Perform CEEMDAN decomposition: Among them, IMF k,j (t) is the intrinsic modal component decomposed by CEEMDAN, J is the decomposition order, r k (t) is the residual term; CEEMDAN avoids the modal aliasing problem in the traditional EMD method by adding adaptive noise, making the decomposition result more stable; Step 2.

3. Optimize feature fusion with dynamic weights Finally, we get the set of subsequences after multi-scale decomposition: D={IMF k,j (t)|k=1,2,...,K;j=1,2,...,J} Different IMF k,j (t) has different degrees of influence on the target variable. In order to optimize data fusion, the dynamic grey correlation degree is used again to calculate the weight of each subsequence and construct the final fusion feature data set: in, represents the final fused dataset combined with dynamic weights, W t represents the normalized dynamic weight, which indicates the relative influence of the j-th IMF of the k-th VMD component on the target variable.

4. The short-term load forecasting method for a PV-storage-direct-flexible system based on multi-level data association fusion according to claim 1 is characterized by: The multimodal attention fusion prediction model developed in step 3 can be expressed as follows: Step 3.1 Design a spatiotemporal joint attention mechanism The spatiotemporal joint attention mechanism is used for feature fusion. This mechanism is based on the architecture of the Transformer model and introduces a spatiotemporal joint attention module to capture the long-term dependencies in different modalities and time sequences. The spatiotemporal joint attention mechanism performs weighted summation of the multimodal input data at each time step through the self-attention mechanism, thereby effectively focusing on the time step and mode that are most important for prediction, ensuring that the model can weight different input signals according to actual conditions when processing different types of loads, and obtain accurate short-term power load forecasts; Step 3.2 Use a hierarchical distillation strategy to optimize the computational efficiency of the self-attention module The distillation strategy reduces the complexity of attention calculation layer by layer, making the high-level feature representation more concise and efficient. It also reduces unnecessary calculations of the model by gradually learning simplified attention patterns, thereby improving the overall computational efficiency. Through layered distillation, unnecessary computation in the self-attention mechanism is reduced, reducing the time cost of model training and inference, so that the system can respond to changing load data in real time; Step 3.3 Introduce a bidirectional GRU network at the encoder layer A bidirectional GRU network is introduced in the encoder layer of Transformer. The bidirectional GRU can consider both forward and backward time series information at the same time, enhancing the model's sensitivity to historical load data and future load changes. The bidirectional GRU network can obtain the time series characteristics of load data from both the forward and backward directions. In this way, the dynamic evolution pattern of load data can be better understood, thereby improving the prediction accuracy of load changes. Step 3.4 Embed the dynamic coupling analysis module at the decoder layer At the decoder level, a dynamic coupling analysis module is added to capture the nonlinear interaction between different data sources; This module can dynamically capture the nonlinear interaction patterns between loads, combine the forward and backward time series information, explicitly model the interaction characteristics between load data, improve the expressiveness of the model, and make the final prediction results more accurate; Step 3.5 Explicitly model long-range correlations With the support of the joint spatiotemporal attention mechanism, the model can explicitly model long-range correlations in high-dimensional data; through the Transformer architecture, especially the introduction of the self-attention mechanism, the model can pay attention to the long-range dependency characteristics across time steps through the global attention mechanism, so that the mutual influence between multiple data sources such as photovoltaics, electric vehicle charging loads and building loads can be fully learned.

Images