Method and system for short-term prediction of user-side electricity consumption power of distributed access power grid
By standardizing and weighting the user-side data of distributed photovoltaic power grid integration, and combining it with deep learning algorithms, feature functions and decomposition models are constructed, solving the problem of low accuracy in power consumption prediction for distributed photovoltaic power grid integration, and achieving higher accuracy prediction and grid optimization.
Patent Information
- Application Number
- CN202511021468.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-24
- Publication Date
- 2025-11-28
AI Technical Summary
Existing technologies for short-term prediction of user-side power consumption after distributed photovoltaic power is connected to the grid suffer from low prediction accuracy and strong dependence on the quality of models, parameters, and basic data.
By performing range scaling and standardization on user-side data of distributed photovoltaic grid-connected power, and then weighting and integrating the standardized data, a feature function and decomposed power prediction model are constructed. Deep learning algorithms are used for data processing to capture the short-term time-series relationship and the influence of complex factors on user-side power consumption, enabling accurate prediction.
It improves the accuracy of user-side power consumption forecasting, reduces reliance on model and data quality, optimizes grid dispatching and operation, enhances power supply reliability and economy, and supports grid planning and the consumption of new energy sources.
Smart Images
Figure CN121035971A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the field of power system load forecasting, in particular to a user-side power consumption short-term forecasting method and system for distributed access to power grids. BACKGROUND
[0002] Under the background of energy transformation and sustainable development, distributed photovoltaic power generation has been expanding in the power grid due to its clean and renewable advantages. After the distributed photovoltaic power generation is connected to the power grid, short-term forecasting of user-side power consumption becomes an important problem to be solved. Due to the influence of light, temperature, cloud and other meteorological factors on photovoltaic output, there is intermittency, volatility and randomness. In addition, user power consumption behavior is influenced by living habits, working hours, seasonal changes and other factors, which are diverse and random. Moreover, power consumption is also affected by the coupling of multiple factors such as power grid operation state and electricity market price, forming a complex nonlinear system, which makes it a great challenge to accurately predict user-side power consumption. Therefore, it is necessary to study the short-term forecasting of user-side power consumption of distributed photovoltaic access to the power grid.
[0003] The current power system uses various methods to forecast user-side power consumption in the short term, but the current forecasting methods have low prediction accuracy and are highly dependent on models, parameters and basic data quality. SUMMARY
[0004] In order to solve the problems of low prediction accuracy and strong dependence on models, parameters and basic data quality of current forecasting methods, the present application proposes a short-term forecasting method for user-side power consumption of distributed access to the power grid, which comprises: The user-side data of the distributed photovoltaic access to the power grid is subjected to range scaling standardization processing, and the standardized data is weighted and integrated to obtain an integrated feature vector. The integrated feature vector is substituted into a pre-constructed power prediction model to obtain a power prediction value.
[0005] Preferably, the step of substituting the integrated feature vector into the pre-constructed power prediction model to obtain the power prediction value comprises substituting the integrated feature vector into a pre-determined characteristic function to obtain the power prediction value. Or substitute the integrated feature vector into a pre-determined decomposition power prediction model to obtain a load prediction value, wherein the decomposition load prediction model is constructed based on historical load data by discrete wavelet decomposition combined with a deep learning algorithm. The power prediction model comprises a characteristic function and a decomposition power prediction model.
[0006] Preferably, the construction of the characteristic function comprises: extracting user-side power features and short-term time series relationship of user-side power from the user-side historical integrated power consumption data; mapping the short-term time series features and power consumption features into a high-dimensional space for feature fitting to construct an initial feature function; substituting the historical integrated power consumption data into the initial feature function to obtain power consumption prediction values; calculating prediction errors based on the power consumption prediction values minus actual power values corresponding to the integrated feature vectors; correcting the initial feature function based on the prediction errors to obtain a feature function.
[0007] Preferably, the extracting user-side power features and short-term time series relationship of user-side power from the user-side historical integrated power consumption data comprises: calculating distributed photovoltaic power features based on the user-side historical integrated data combined with a distributed photovoltaic power feature calculation formula; calculating power consumption features based on the distributed photovoltaic power features combined with a power consumption feature calculation formula; determining short-term time series relationship of user-side power based on multiple user-side power consumption features.
[0008] Preferably, the power consumption feature calculation formula is as follows:
[0009]
[0010] In the formula, represents instantaneous power of the user side, represents real-time voltage of the user side, represents real-time current of the user side, represents power factor of the user side, represents distributed photovoltaic power connected to the power grid, represents average power of the user side, represents sampling period of the user side, represents power fluctuation function of the user side.
[0011] Preferably, the short-term time series relationship of user-side power is as follows:
[0012] In the formula, represents user-side power time series features, represents hysteresis of power consumption, represents power change function of the user side, represents power consumption variance, a sampling period of the user side, a power fluctuation function of the user side, m represents a sampling period sequence number of the user side, a power change function of the user side in the m+k period.
[0013] Preferably, the characteristic function is as follows:
[0014] In the formula, representing the constructed user side power consumption characteristic function, representing a mapping parameter, representing a characteristic mapping function, , respectively representing weight values of the influence degree of different characteristics on power consumption, representing a bias term, representing a user side power consumption time sequence characteristic, representing the instantaneous power of the user side, representing the average power of the user side, representing the load rate of the user side power consumption, representing the peak-valley difference of the user side power consumption.
[0015] Preferably, the mapping of the short-term time sequence characteristic and the power consumption characteristic into a high-dimensional space for characteristic fitting to construct an initial characteristic function comprises: constructing a sample set by taking the characteristics affecting power consumption as input characteristics and taking the historical power consumption of the user side as output characteristics; selecting a Gaussian kernel function as a characteristic mapping function, and performing parameter optimization on the kernel function parameters on the sample set through cross-validation to obtain optimal kernel function parameters; mapping the input characteristics of the sample set into a high-dimensional space through the Gaussian kernel function to construct a linear regression model; converting the linear regression model into a dual form, and solving a dual problem based on the input characteristics and the output characteristics of the sample set to obtain prediction coefficients; obtaining an initial characteristic function from the prediction coefficients combined with the Gaussian kernel function.
[0016] Preferably, the substituting the integrated characteristic vector into the pre-determined decomposition power prediction model to obtain a power prediction value comprises: adopting discrete wavelet transform to decompose the integrated characteristic vector into a high-frequency abrupt change component and a low-frequency periodic component; adopting a gated recurrent unit to capture time sequence dependence in the high-frequency abrupt change component, and combining a gated linear unit to enhance nonlinear expression capability; The Transformer encoder is used to learn the global dependency of the mode in the low-frequency periodic component, and the time synchronization information is reserved through the position encoding; The weights of the high-frequency mutation component and the low-frequency periodic component are adaptively adjusted according to the mutation intensity at the current moment through the learnable weighting coefficients. The power prediction value is obtained from the high-frequency mutation component, the low-frequency periodic component and the respective weights.
[0017] Preferably, the power prediction value is calculated as follows:
[0018] In the formula, represents the prediction kernel function of the cth iteration, represents the prediction function of the user-side power consumption, represents the prediction coefficient, represents the number of iterations, represents the prediction error value of the user side.
[0019] Preferably, the weighted integration of the normalized data comprises: The user-side power consumption data is divided into nodes at the "user-equipment" level, with the user as the parent node and the equipment as the child node, and multiple users in the same area form a regional node. Edges are set between the regional nodes, parent nodes and child nodes. The correlation strength between the nodes is calculated based on the historical power consumption data, and the correlation strength is used as the edge weight. An improved attention mechanism is used to assign adaptive attention weights to each node. The power consumption data on each node is combined with the adaptive attention weights to perform weighted integration, and the integrated data is obtained.
[0020] In another aspect, the application also discloses a distributed user-side power consumption short-term prediction system for accessing the power grid, comprising: A preprocessing module is configured to perform range scaling normalization processing on the user-side data of the distributed photovoltaic access power grid, and to perform weighted integration on the normalized data to obtain an integrated feature vector. A power prediction module is configured to substitute the integrated feature vector into a pre-constructed power prediction model to obtain a power prediction value.
[0021] Preferably, the power prediction module comprises: A function solving submodule is configured to substitute the integrated feature vector into a pre-determined feature function to obtain a power prediction value. Or a decomposition solving submodule is configured to substitute the integrated feature vector into a pre-determined decomposition power prediction model to obtain a consistent prediction value, wherein the decomposition load prediction model is constructed based on the historical load data by discrete wavelet decomposition and combined with a deep learning algorithm. The power prediction model comprises a characteristic function and a decomposed power prediction model.
[0022] Preferably, the characteristic function construction module is configured to: extract power characteristics of user-side power consumption and short-term time-series relationship of user-side power from historical user-side power consumption data; map the short-term time-series characteristics and power consumption characteristics to a high-dimensional space for characteristic fitting to construct an initial characteristic function; substitute historical user-side power consumption data into the initial characteristic function to obtain power consumption prediction values; calculate prediction errors based on the power consumption prediction values minus actual power values corresponding to the integrated characteristic vectors; correct the initial characteristic function based on the prediction errors to obtain the characteristic function.
[0023] Preferably, the characteristic function construction module extracts power characteristics of user-side power consumption and short-term time-series relationship of user-side power from historical user-side power consumption data, and the specific implementation comprises: calculate distributed photovoltaic power characteristics based on historical user-side integrated data and a distributed photovoltaic power characteristic calculation formula; calculate power consumption characteristics based on the distributed photovoltaic power characteristics and a power consumption characteristic calculation formula; determine short-term time-series relationship of user-side power based on multiple user-side power consumption characteristics.
[0024] Preferably, the power consumption characteristic calculation formula is as follows:
[0025]
[0026] In the formula, represents instantaneous power of the user side, represents real-time voltage of the user side, represents real-time current of the user side, represents power factor of the user side, represents distributed photovoltaic power connected to the power grid, represents average power of the user side, represents sampling period of the user side, represents power fluctuation function of the user side.
[0027] Preferably, the short-term time-series relationship of the user-side power is as follows:
[0028] In the formula, representing the time sequence characteristics of the power consumption on the user side, representing the hysteresis of the power consumption, representing the power change function on the user side, representing the variance of the power consumption, representing the sampling period on the user side, representing the power fluctuation function on the user side, m represents the sampling period sequence number on the user side, representing the power change function on the user side in the m+k period.
[0029] Preferably, the characteristic function is as follows:
[0030] In the formula, representing the constructed power consumption characteristic function on the user side, representing the mapping parameter, representing the characteristic mapping function, , respectively representing the weight value of the influence degree of different characteristics on the power consumption, representing the bias term, representing the time sequence characteristics of the power consumption on the user side, representing the instantaneous power on the user side, representing the average power on the user side, representing the load rate of the power consumption on the user side, representing the peak-valley difference of the power consumption on the user side.
[0031] Preferably, the specific implementation steps of constructing the initial characteristic function by mapping the short-term time sequence characteristics and the power consumption characteristics into a high-dimensional space for characteristic fitting in the characteristic function construction module include: constructing a sample set by taking the characteristics affecting the power consumption as input characteristics and taking the historical power consumption on the user side as output characteristics; selecting a Gaussian kernel function as a characteristic mapping function, and performing parameter optimization on the kernel function parameters on the sample set through cross-validation to obtain optimal kernel function parameters; mapping the input characteristics of the sample set into a high-dimensional space through the Gaussian kernel function to construct a linear regression model; converting the linear regression model into a dual form, solving a dual problem based on the input characteristics and the output characteristics of the sample set to obtain prediction coefficients; obtaining the initial characteristic function from the prediction coefficients combined with the Gaussian kernel function.
[0032] Preferably, the power prediction module is specifically used for: adopting discrete wavelet transform to decompose the integrated characteristic vector into a high-frequency abrupt change component and a low-frequency periodic component; The gating cycle unit is adopted to capture the time dependence in the high-frequency mutation component, and the gating linear unit is adopted to enhance the nonlinear expression capability; The Transformer encoder is adopted to learn the global dependence of the mode in the low-frequency periodic component, and the position encoding is used to reserve the time synchronization information; The weight of the high-frequency mutation component and the low-frequency periodic component is adaptively adjusted according to the mutation intensity at the current moment through the learnable weighting coefficient; The power prediction value is obtained from the high-frequency mutation component, the low-frequency periodic component and the respective corresponding weight.
[0033] Preferably, the power prediction value is calculated as follows:
[0034] In the formula, represents the prediction kernel function of the cth iteration, represents the prediction function of the user-side power consumption, represents the prediction coefficient, represents the number of iterations, represents the prediction error value of the user side.
[0035] Preferably, the data after the standardization processing is weighted and integrated in the preprocessing module, and the specific implementation steps include: The user-side power consumption data is divided into nodes according to the "user-equipment" level, the user is taken as a parent node, the equipment is taken as a child node, multiple users in the same area form an area node, and edges are set between the area node, the parent node and the child node; The association strength between the nodes is calculated based on the historical power consumption data, and the association strength is taken as the edge weight; An improved attention mechanism is adopted to assign an adaptive attention weight to each node, and the power consumption data on each node is weighted and integrated to obtain the integrated data in combination with the adaptive attention weight.
[0036] In another aspect, the present application also provides an electronic device, comprising: at least one processor and a memory; the memory and the processor are connected through a bus; The memory is used to store one or more programs; When the one or more programs are executed by the at least one processor, a distributed access power grid user-side power consumption short-term prediction method as described above is implemented.
[0037] In another aspect, the present application also provides a readable storage medium, which has an execution program stored thereon, and the execution program is executed to implement a distributed access power grid user-side power consumption short-term prediction method as described above.
[0038] Compared with the prior art, the beneficial effects of this application are as follows: This application provides a short-term prediction method for user-side power consumption in distributed photovoltaic grid-connected systems. The method includes: performing range scaling and standardization on user-side power consumption data from distributed photovoltaic grid-connected systems, and then weighting and integrating the standardized data; finally, substituting the integrated feature vector into a pre-constructed power prediction model to obtain the predicted power value. This application performs range scaling and standardization on user-side power consumption data and then weights and integrates the standardized data, thus consolidating scattered data, eliminating data inconsistencies and fragmentation, forming a complete and unified dataset, and improving data quality. This application also utilizes feature functions constructed from the temporal relationships between different power consumption characteristics on the user side to perform power prediction, thereby improving prediction accuracy and eliminating dependence on the quality of models, parameters, and basic data. Attached Figure Description
[0039] Figure 1 This is a flowchart of the short-term power consumption prediction method for the user side of the distributed grid access in this application; Figure 2 This is a schematic diagram of an electronic device structure according to this application. Detailed Implementation
[0040] Traditional methods treat user-side power characteristics as a whole, neglecting the analysis of the time-series relationships between different power consumption characteristics. This fails to capture the changing patterns of photovoltaic output and user electricity consumption behavior, resulting in significant deviations between power predictions and actual conditions.
[0041] This application proposes a short-term forecasting method for user-side power consumption in a distributed grid, which can significantly improve the intelligence level of grid dispatching and operation, optimize power resource allocation, reduce grid imbalance, improve power supply reliability and economy, and also provide a reference for the planning and design of new energy power plants, promoting the consumption and utilization of new energy.
[0042] To better understand this application, the content of this application will be further described below in conjunction with the accompanying drawings and embodiments.
[0043] Example 1: A method for short-term prediction of user-side power consumption in a distributed grid-connected system, such as... Figure 1 As shown, it includes: Step 1: Perform range scaling and standardization on the user-side data of distributed photovoltaic grid connection, and then weight and integrate the standardized data to obtain the integrated feature vector; Step 2: Substitute the integrated feature vector into the pre-built power prediction model to obtain the power prediction value.
[0044] The following is a further description of each step in this application: Step 1: Perform range scaling and standardization on the user-side data of distributed photovoltaic grid connection, and then weight and integrate the standardized data to obtain an integrated feature vector, including: The user-side electricity consumption data is normalized using a standardized phase formula; The data after planning is integrated using an integration phase formula.
[0045] Furthermore, the formula for the normalization stage is as follows:
[0046] In the formula, This indicates the result of user-side data normalization processing. This represents the raw user-side data. This represents the mean of the user-side data. This represents the standard deviation of the user-side data. This represents the maximum value of the user-side data. This represents the minimum value of the user-side data. Represents the normalization coefficient.
[0047] Furthermore, the formula for the integration stage is as follows:
[0048] In the formula, This represents the integrated user-side data. Indicates the integrated parameters. This represents the integrated weight value of data from different user sides. This indicates the amount of user-side data that needs to be integrated, and i represents the user-side sequence number that needs to be integrated.
[0049] Step 1 is as follows: I. Integration of User-Side Electricity Consumption Data for Distributed Photovoltaic Power Grid Connection User-side electricity consumption data comes from a wide range of sources and is complex, involving many different types of users with significantly different electricity consumption behaviors and habits. The data may be scattered across different systems or devices, and its formats and standards may vary. Furthermore, the integration of distributed photovoltaic (PV) power means that electricity consumption is affected by a combination of factors, including PV output and weather conditions. Only by integrating this scattered data can inconsistencies and fragmentation be eliminated, forming a complete and unified dataset. Therefore, the first step is to integrate the user-side electricity consumption data from distributed PV grid connections to improve data quality.
[0050] In this process, the collected data is first standardized. The specific formula for the standardization stage is shown below: (1) In the formula, This indicates the result of user-side data normalization processing. This represents the raw user-side data. This represents the mean of the user-side data. This represents the standard deviation of the user-side data. This represents the maximum value of the user-side data. This represents the minimum value of the user-side data. Represents the normalization coefficient.
[0051] Based on the normalized data, an integration phase formula is used to integrate user-side data and distributed photovoltaic power generation data. The specific integration phase formula is shown below: (2) In the formula, This represents the integrated user-side data. Indicates the integrated parameters. This represents the integrated weight value of data from different user sides. This indicates the amount of user-side data that needs to be integrated, and i represents the user-side sequence number that needs to be integrated.
[0052] Based on the above formula, the user-side data is integrated and processed, laying the foundation for the subsequent construction of the user-side power consumption characteristic equation.
[0053] Step 2: Substitute the integrated feature vector into the pre-built power prediction model to obtain the power prediction value, including: Substituting the integrated feature vector into a pre-determined feature function yields the predicted power value; Alternatively, the integrated feature vector can be substituted into a pre-determined decomposed power prediction model to obtain the load prediction value. The decomposed load prediction model is constructed based on discrete wavelet decomposition of historical load data and combined with deep learning algorithms. The power prediction model includes a characteristic function and a decomposed power prediction model.
[0054] The characteristic function and the decomposed power prediction model are introduced below: (1) The construction of the characteristic function includes: From the historical integrated user-side electricity consumption data, extract user-side power characteristics and short-term time-series relationships of user-side power; The short-term time-series features and power consumption features are mapped to a high-dimensional space for feature fitting to construct an initial feature function; Substituting the historically integrated electricity consumption data into the initial characteristic function yields the predicted power consumption value. The prediction error is calculated by subtracting the actual power value corresponding to the integrated feature vector from the predicted power consumption value. The initial feature function is corrected based on the prediction error to obtain the feature function.
[0055] Furthermore, the step of mapping the short-term time-series features and power consumption features to a high-dimensional space for feature fitting and constructing an initial feature function includes: A sample set is constructed by using the features that affect power consumption as input features and the historical power consumption on the user side as output features. The Gaussian kernel function is selected as the feature mapping function, and the kernel function parameters are tuned on the sample set through cross-validation to obtain the optimal kernel function parameters. A linear regression model is constructed by mapping the input features of the sample set to a high-dimensional space using a Gaussian kernel function. The linear regression model is converted into its dual form, and the dual problem is solved based on the input and output features of the sample set to obtain the prediction coefficients. The characteristic function is obtained by combining the prediction coefficients with the Gaussian kernel function.
[0056] The Gaussian kernel function is shown in the following formula: (13) In the formula, Indicates the prediction kernel function, Indicates predictor factor, This represents the mapping parameters used in power prediction. This represents the standard deviation of power.
[0057] The construction of the feature function specifically includes: User-side electricity consumption is influenced by a variety of complex factors, such as different users' lifestyles, work patterns, seasonal changes, and holidays. These factors lead to diversity and randomness in electricity consumption behavior. By analyzing electricity consumption characteristics, we can gain a deeper understanding of the inherent patterns and trends in user electricity consumption behavior, such as peak and off-peak periods, and the correlation between electricity consumption and specific factors. Therefore, based on the integrated user-side electricity consumption data mentioned above, we analyze the original power characteristics of distributed photovoltaic (PV) power, and then calculate the user-side electricity consumption characteristics under the influence of these power characteristics. In this process, the distributed PV power characteristics are calculated as follows: (3) (4) In the formula, This indicates the distributed photovoltaic power generation capacity connected to the power grid. Indicates the intensity of light received by the photovoltaic module. This indicates the area of the distributed photovoltaic modules. This indicates the conversion efficiency of distributed photovoltaic modules. This indicates the power generation efficiency of distributed photovoltaic modules. This represents the actual power generation after integrating user-side data and distributed photovoltaic power generation data. This represents the theoretical power generation of a photovoltaic module.
[0058] Based on the aforementioned characteristics of distributed photovoltaic power, the power consumption characteristics on the user side are extracted. The extracted basic power characteristics are shown below: (5) (6) In the formula, This represents the instantaneous power on the user side. This represents the real-time voltage on the user side. This represents the real-time current on the user side. This represents the power factor on the user side. This indicates the distributed photovoltaic power generation capacity connected to the power grid. This represents the average power on the user side. Indicates the sampling period on the user side. This represents the power fluctuation function on the user side.
[0059] Based on the above formula, multiple power features are extracted in depth. The details are as follows: (7) (8) In the formula, This indicates the load factor of the user-side power consumption. This represents the peak-to-valley difference in power consumption on the user side. This indicates the maximum power consumption on the user side. This indicates the minimum power consumption on the user side.
[0060] Based on the extracted power features, corresponding characteristic equations are constructed, laying the foundation for subsequent prediction of user-side power consumption. Furthermore, based on the extracted user-side power consumption features, short-term time-series relationships of user-side power are captured, thereby enabling short-term power consumption prediction. The specific process is as follows: (9) In the formula, This indicates the timing characteristics of power consumption on the user side. This indicates the lag in power consumption. This represents the power change function on the user side. This represents the variance of electrical power consumption. Indicates the sampling period on the user side. This represents the power fluctuation function on the user side, where m represents the user-side sampling period number. This represents the power change function on the user side during the m+k period.
[0061] Based on the short-term time-series relationship of the user-side power, and combined with the power consumption features extracted by formulas (5)-(8), they are mapped to a high-dimensional space and feature-fitted to construct the corresponding initial feature function as shown in formula (10): (10) In the formula, This represents the constructed user-side power consumption characteristic function. Indicates the mapping parameters. Represents the feature mapping function. , These represent the weight values for the degree of influence of different characteristics on power consumption. Indicates the bias term. This indicates the timing characteristics of power consumption on the user side. This represents the instantaneous power on the user side. This represents the average power on the user side. This indicates the load factor of the user-side power consumption. This represents the peak-to-valley difference in power consumption on the user side.
[0062] Based on the constructed feature function, power consumption is predicted, and the corresponding prediction error is calculated. The specific calculation formula is shown below: (11) (12) In the formula, This represents the loss function on the user side in predicting power consumption. Indicates the loss coefficient. This represents the error generalization factor during prediction. Let N represent the fitting error, N represent the number of iterations, and c represent the iteration number. This represents the prediction error value on the user side. This represents the penalty factor.
[0063] The prediction error is used as a correction term to correct the initial characteristic function, resulting in a characteristic function. The user-side power consumption is then predicted using this characteristic function, including: The predicted power value is calculated using the following formula: (14) In the formula, Indicates the prediction kernel function, Indicates predictor factor, This represents the mapping parameters used in power prediction. Indicates the standard deviation of power. The prediction function representing the user-side power consumption. Represents the prediction coefficient. Indicates the number of iterations.
[0064] This application utilizes advanced data mining technology to conduct in-depth analysis of a large amount of historical electricity consumption data and distributed power generation data, uncovering the hidden short-term time-series relationships of user-side power, and providing support for more accurate prediction.
[0065] (2) The construction of the decomposed power prediction model includes: Discrete wavelet transform is used to decompose the integrated eigenvector into decomposition modules of high-frequency abrupt change components and low-frequency periodic components; A high-frequency mutation component processing module that uses gated loop units to capture the temporal dependencies in high-frequency mutation components and then outputs the prediction results of high-frequency mutation components; A low-frequency periodic component processing module is used to process low-frequency periodic components with a ransformer encoder and capture mode dependencies in long sequences through a self-attention mechanism, thereby outputting the prediction results of low-frequency periodic components. An adaptive weight fusion module that performs adaptive weight fusion on the prediction results of high-frequency mutation components and low-frequency periodic components; The decomposed power prediction model is constructed by a decomposition module, a high-frequency mutation component processing module, a low-frequency periodic component processing module, and an adaptive weight fusion module.
[0066] Furthermore, by substituting the integrated feature vector into the pre-determined decomposed power prediction model, the power prediction values are obtained, including: Step 21: By using the decomposition module in the power prediction model, the integrated feature vector is decomposed into high-frequency abrupt change components and low-frequency periodic components using discrete wavelet transform. Step 22: By using a gated cyclic unit to capture the temporal dependence of the high-frequency mutation component in the high-frequency mutation component processing module in the decomposition power prediction model, and combining it with a gated linear unit to enhance the nonlinear expression capability; Step 23: By using the low-frequency periodic component processing module in the decomposed power prediction model, the Transformer encoder is used to learn the global dependency of the mode in the low-frequency periodic component, and the time synchronization information is preserved by position encoding. Step 24: By using the adaptive weight fusion module in the decomposed power prediction model, the weights of the high-frequency mutation component and the low-frequency periodic component are adaptively adjusted according to the mutation intensity at the current moment through learnable weighting coefficients to obtain the power prediction value.
[0067] Step 25: Based on the decomposed power prediction model, predict the power consumption on the user side to ensure the stable operation of the power system.
[0068] The following further clarifies each step: Step 21 specifically includes: When processing the feature vector, the integrated feature vector is first decomposed into high-frequency abrupt change components and low-frequency periodic components through discrete wavelet transform (DWT) to extract feature data of different frequencies, laying the foundation for subsequent processing.
[0069] Step 22 specifically includes: For the processing of high-frequency mutation components, we first assume that the data obtained from the previous stage is a time series data containing high-frequency mutations, such as the mutation frequency data of a gene sequence within a specific time window, with a data length of 1000 time points, a mutation frequency value between 0 and 1 at each point, and a sampling interval of 1 hour.
[0070] In the initial data, high-frequency components were extracted. It was assumed that, through Fast Fourier Transform (FFT) analysis, the portion with frequencies above 0.1 Hz was identified as high-frequency abrupt change components, accounting for approximately 30% of the total signal, or 300 valid data points, with an average abrupt change frequency of 0.65. To capture the temporal dependencies in these high-frequency abrupt change components, a Gated Recurrent Unit (GRU) model was applied. The GRU parameters were set as follows: hidden layer dimension 64, input layer dimension 1 (corresponding to a single abrupt change frequency value), output layer dimension 1, 50 training epochs, a learning rate of 0.001, and mean squared error (MSE) as the loss function.
[0071] During training, the data was divided into an 80% training set (240 data points) and a 20% test set (60 data points). Temporal inputs were constructed using a sliding window (window size 10) to ensure that the model could learn short-term dependencies.
[0072] After training, the model achieved an MSE value of 0.012 on the test set, indicating that the model effectively captured the short-term dynamic trends in high-frequency mutation components.
[0073] Ultimately, the GRU model outputs the predicted mutation frequency value for each time point, forming a dynamic feature sequence of length 300, which reflects the fluctuation pattern of mutation frequency in the short term. For example, at time points 50 to 60, the mutation frequency increases from 0.62 to 0.68, showing a clear short-term upward trend.
[0074] These dynamic features will be used as input for subsequent fusion processing, supporting the further integration of low-frequency components and other biological information. In this way, a complete workflow from high-frequency mutation components to dynamic feature extraction is achieved, ensuring that the temporal relevance of information is preserved, while providing a reliable data foundation for subsequent analysis.
[0075] Step 23 specifically includes: For processing low-frequency periodic components, the low-frequency components are first extracted from the input signal. Assume that the input signal is a time series data with a sampling frequency of 100Hz and contains 1000 sampling points.
[0076] The signal is analyzed in the frequency domain by using the Fast Fourier Transform (FFT) algorithm. The frequency threshold is set to 5Hz, and the components below this threshold are extracted as low-frequency periodic components.
[0077] The calculation results show that the low-frequency components are mainly concentrated in the range of 0.5Hz to 2Hz, accounting for about 60% of the total signal energy, indicating that they contain significant periodic information.
[0078] Next, a transformer encoder is applied to learn the global dependency pattern of low-frequency components. Specifically, the extracted low-frequency sequence is divided into subsequences of length 50. Each subsequence is input into a model containing a 6-layer transformer encoder with 8 attention heads per layer and a hidden layer dimension of 512. The long-term dependency relationship between time points within the sequence is captured through a multi-head attention mechanism. The mean squared error loss function is used during training, the learning rate is set to 0.001, and the model is iterated for 100 rounds. After the model converges, the global dependency feature matrix is output.
[0079] Subsequently, to preserve the sequence order information, position encoding is introduced. Sine and cosine functions are used to generate position vectors with dimensions consistent with the input sequence. For example, for the i-th position, the position encoding values are sin(i / 10000^(2j / d)) and cos(i / 10000^(2j / d)), where j is the dimension index and d is the total dimension of 512, ensuring that the temporal characteristics of the sequence are not lost.
[0080] Finally, through the above processing, the long-term regularity pattern in the low-frequency periodic characteristics is obtained. The analysis shows that its periodicity is mainly manifested as a complete cycle of 200 sampling points, which is consistent with the daily traffic fluctuation cycle in actual business scenarios, providing periodic information support for subsequent fusion processing.
[0081] Step 24 specifically includes: Step S1: Calculate the short-term dynamic trend of high-frequency mutation characteristics, calculate the magnitude of change, and use the standard deviation calculation method to quantify the magnitude of change to determine the fluctuation index data of high-frequency mutation characteristics.
[0082] Step S2: Based on the fluctuation index data obtained in Step S1, and combined with the long-term regularity pattern of low-frequency cycle characteristics, preliminary fusion is performed using preset weighting coefficients. If the fluctuation index data exceeds the preset threshold range, the weight of the high-frequency mutation characteristics is increased to obtain the preliminary adjusted weight allocation scheme.
[0083] Step S3: For the preliminary weight allocation scheme obtained in step S2, an adaptive adjustment mechanism is adopted, and the weighting coefficients are optimized and iterated through the gradient descent algorithm to obtain the optimized weight coefficient data.
[0084] Step S4: Based on the optimized weight coefficient data obtained in step S3, perform weighted fusion processing on the high-frequency mutation features and the low-frequency periodic features to generate comprehensive feature representation data.
[0085] Step S5: For the comprehensive feature representation data generated in step S4, the feature importance is evaluated using a preset classification model. If the contribution of a certain feature is lower than a preset threshold, its weight is fine-tuned a second time to determine the final feature fusion result.
[0086] Step S6: Based on the final feature fusion result determined in step S5, generate feature vector data for subsequent analysis modules to call, thus completing the closed-loop process of feature processing.
[0087] When processing time series data to predict the future trend of a certain indicator, the first step is to analyze the short-term dynamic change trend of high-frequency mutation characteristics. For example, if the vibration frequency data of an industrial equipment is collected every minute, the time span is the past 24 hours, and there are a total of 1440 data points.
[0088] Using the sliding window method, with a window size of 10 minutes, the mean and rate of change of vibration frequency within each window were calculated. It was found that the mean rate of change of vibration frequency in the most recent hour was 5.2%, and it showed an accelerating upward trend, indicating the presence of high-frequency abrupt change characteristics.
[0089] Next, for the long-term regularity pattern extraction of low-frequency periodic characteristics, Fourier transform was used to perform frequency domain analysis on the data of the past 30 days, which identified a significant 24-hour period in the vibration frequency, with the peak occurring at 8 am every day and an amplitude of 3.8 units, indicating that the operation of the equipment is affected by the daily working cycle.
[0090] Subsequently, an adaptive weighted adjustment was performed based on the two features mentioned above. For the change amplitude of the high-frequency abrupt change feature, the standard deviation of the vibration frequency in the most recent hour was calculated to be 2.5, reflecting its large fluctuation. The standard deviation of the low-frequency periodic feature was 1.2, indicating relatively stable fluctuation. Therefore, the weights were dynamically allocated by the ratio of the standard deviations, and the weights of the high-frequency feature were calculated to be 2.5 / (2.5+1.2)=0.68 and the weights of the low-frequency feature were calculated to be 1.2 / (2.5+1.2)=0.32.
[0091] Finally, a prediction model is constructed using the weighted eigenvalues. Assuming the current vibration frequency is 10.5 units, the prediction increment for high-frequency features is 0.7, and the prediction increment for low-frequency features is 0.3, then the weighted prediction increment is 0.7×0.68+0.3×0.32=0.572, and the predicted vibration frequency for the next moment is 10.5+0.572=11.072 units.
[0092] Using the above methods, the system can automatically adjust the weights according to the importance of features to achieve accurate prediction of equipment status. At the same time, combined with the equipment maintenance needs in industrial scenarios, if the predicted value exceeds the threshold of 11.0 units, the system will automatically generate an early warning signal to ensure the safe operation of the equipment.
[0093] The short-term power consumption forecasting method for user-side distributed grid access provided in this application has the following effects: I. High prediction accuracy 1. This method can comprehensively consider various factors affecting power consumption, such as the output characteristics of distributed power sources, the power consumption behavior patterns of users, and meteorological factors. By establishing complex prediction models and algorithms, it can capture the complex relationship between these factors and power consumption, thereby improving the accuracy of prediction.
[0094] 2. This method utilizes advanced data mining techniques to conduct in-depth analysis of a large amount of historical electricity consumption data and distributed power source data, uncovering hidden patterns and trends to support more accurate predictions.
[0095] II. Effectively improve the stability of power grid operation 1. It helps grid dispatchers understand the electricity demand on the user side and the output of distributed power sources in advance, so as to arrange the grid operation mode and dispatch plan more rationally, optimize the allocation of power resources, reduce grid load fluctuations, and improve the stability and reliability of grid operation.
[0096] 2. It can detect potential problems and risks in power grid operation in advance, such as local overload and voltage abnormality, and take timely measures to adjust and prevent them, thereby reducing the probability of power grid failure and ensuring the safe operation of the power grid.
[0097] III. Strong support for power grid planning and decision-making 1. It provides important reference for the planning and construction of the power grid. By forecasting the power consumption on the user side in the short term, it can help understand the load growth trend and distribution of the power grid in the future, and help power grid planners to rationally plan the layout and capacity of the power grid, thereby improving the scientificity and rationality of power grid planning.
[0098] 2. It can be used to assess the impact of different power policies and measures on user-side power consumption, provide data support for the formulation and adjustment of power policies, and promote the healthy development of the power market.
[0099] Example 2: Based on the same inventive concept, this application also provides a short-term power consumption forecasting system for the user side of a distributed grid-connected system, comprising: The preprocessing module is used to perform range scaling and standardization on user-side data of distributed photovoltaic grid connection, and to weight and integrate the standardized data to obtain an integrated feature vector. The power prediction module is used to input the integrated feature vector into the pre-built power prediction model to obtain the power prediction value.
[0100] Preferably, the power prediction module includes: The function solving submodule is used to substitute the integrated feature vector into a pre-determined feature function to obtain the power prediction value; Alternatively, a decomposition and solution submodule can be used to substitute the integrated feature vector into a pre-determined decomposition power prediction model to obtain the predicted value. The decomposition load prediction model is constructed based on discrete wavelet decomposition of historical load data and combined with deep learning algorithms.
[0101] Preferably, it further includes: a feature function construction module, used for: An initial feature function is constructed by mapping the short-term time-series features and power consumption features to a high-dimensional space for feature fitting. Based on the initial feature function, historical power consumption is predicted to obtain a predicted power consumption value; the prediction error is calculated by subtracting the actual power value corresponding to the integrated feature vector from the predicted power consumption value; the initial feature function is corrected based on the prediction error to obtain a feature function.
[0102] Preferably, the predicted power value is calculated using the following formula:
[0103] In the formula, This represents the prediction kernel function for the c-th iteration. The prediction function representing the user-side power consumption. Represents the prediction coefficient. Indicates the number of iterations. This represents the prediction error value on the user side.
[0104] Preferably, it further includes a model building module, used for: Discrete wavelet transform is used to decompose the integrated eigenvector into decomposition modules of high-frequency abrupt change components and low-frequency periodic components; A high-frequency mutation component processing module that uses gated loop units to capture the temporal dependencies in high-frequency mutation components and then outputs the prediction results of high-frequency mutation components; A low-frequency periodic component processing module is used to process low-frequency periodic components with a ransformer encoder and capture mode dependencies in long sequences through a self-attention mechanism, thereby outputting the prediction results of low-frequency periodic components. An adaptive weight fusion module that performs adaptive weight fusion on the prediction results of high-frequency mutation components and low-frequency periodic components; The decomposed power prediction model is constructed by a decomposition module, a high-frequency mutation component processing module, a low-frequency periodic component processing module, and an adaptive weight fusion module.
[0105] Preferably, the decomposition and solution submodule is specifically used for: By using the decomposition module in the power prediction model, the integrated eigenvector is decomposed into high-frequency abrupt components and low-frequency periodic components using discrete wavelet transform. By decomposing the high-frequency mutation component processing module in the power prediction model, gated cyclic units are used to capture the temporal dependence of the high-frequency mutation components, and gated linear units are combined to enhance the nonlinear expression capability. By using the low-frequency periodic component processing module in the decomposed power prediction model, a Transformer encoder is employed to learn the global dependency of the mode in the low-frequency periodic component, and time synchronization information is preserved through position encoding. By using the adaptive weight fusion module in the decomposed power prediction model, the weights of high-frequency mutation components and low-frequency periodic components are adaptively adjusted according to the mutation intensity at the current moment through learnable weighting coefficients, thus obtaining the power prediction value.
[0106] This application provides a short-term power consumption prediction system for user-side distributed grid access, which analyzes the time-series relationship between different power consumption characteristics, captures the changing patterns of photovoltaic output and user power consumption behavior, making the predicted power closer to the actual situation, and eliminating the dependence on the quality of models, parameters, and basic data.
[0107] Example 3 like Figure 2As shown, this application also provides an electronic device, which may be a computer device, a microcontroller device, a smart mobile device, etc. The electronic device in this embodiment may include a processor, a memory, a transceiver component, etc. The memory, processor, and transceiver component are connected via a bus; the memory can be used to store executable programs, and an exemplary executable program may include instructions; the processor is used to execute the instructions stored in the memory. The memory can also be used to store data, which can be accessed and / or modified when instructions are executed.
[0108] The processor may be a Central Processing Unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the terminal, and it is suitable for implementing one or more instructions. Specifically, it is suitable for loading and executing one or more instructions in the storage medium to implement the corresponding method flow or corresponding function, so as to realize the steps of the short-term power consumption prediction method for user side of a distributed grid access in the above embodiments.
[0109] Example 4 Based on the same concept, this application also provides a readable storage medium, specifically an electronic device readable storage medium (Memory). This readable storage medium is a memory device within an electronic device used to store programs and data. It is understood that the storage medium here can include both built-in storage media within the electronic device and extended storage media supported by the electronic device. The storage medium provides storage space, which stores the terminal's operating system. Furthermore, this storage space also stores one or more instructions suitable for loading and execution by a processor. These instructions can be one or more executable programs (including program code). It should be noted that the storage medium here can be high-speed RAM or non-volatile memory, such as at least one disk storage device. Loading and executing one or more instructions stored in the storage medium by the processor can implement the steps of the short-term power consumption prediction method for distributed grid access in the above embodiments.
[0110] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0111] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0112] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0113] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0114] The above are merely embodiments of this application and are not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application shall be included within the scope of the claims of this application pending approval.
Claims
1. A method for short-term prediction of user-side power consumption in a distributed grid connection, characterized in that, include: The user-side data of distributed photovoltaic grid connection is subjected to range scaling and standardization processing, and the standardized data is weighted and integrated to obtain the integrated feature vector; The integrated feature vector is substituted into the pre-built power prediction model to obtain the power prediction value.
2. The method as described in claim 1, characterized in that, The step of substituting the integrated feature vector into the pre-built power prediction model to obtain the power prediction value includes: Substituting the integrated feature vector into a pre-determined feature function yields the predicted power value; Alternatively, the integrated feature vector can be substituted into a pre-determined decomposed power prediction model to obtain the load prediction value. The decomposed load prediction model is constructed based on discrete wavelet decomposition of historical load data and combined with deep learning algorithms. The power prediction model includes a characteristic function and a decomposed power prediction model.
3. The method as described in claim 2, characterized in that, The construction of the feature function includes: From the historical integrated user-side electricity consumption data, extract user-side power characteristics and short-term time-series relationships of user-side power; The short-term time-series features and power consumption features are mapped to a high-dimensional space for feature fitting to construct an initial feature function; Substituting the historically integrated electricity consumption data into the initial characteristic function yields the predicted power consumption value. The prediction error is calculated by subtracting the actual power value corresponding to the integrated feature vector from the predicted power consumption value. The initial feature function is corrected based on the prediction error to obtain the feature function.
4. The method as described in claim 3, characterized in that, The extraction of user-side power characteristics and short-term time-series relationships from historically integrated user-side electricity consumption data includes: Based on the historical data integrated from the user side and combined with the distributed photovoltaic power characteristic calculation formula, the distributed photovoltaic power characteristics are calculated. The power consumption characteristics are calculated based on the distributed photovoltaic power characteristics combined with the power consumption characteristics. Based on the power consumption characteristics of multiple users, the short-term time-series relationship of user-side power is determined.
5. The method as described in claim 2, characterized in that, The characteristic function is shown in the following equation: In the formula, This represents the constructed user-side power consumption characteristic function. Indicates the mapping parameters. Represents the feature mapping function. , These represent the weight values for the degree of influence of different characteristics on power consumption. Indicates the bias term. This indicates the timing characteristics of power consumption on the user side. This represents the instantaneous power on the user side. This represents the average power on the user side. This indicates the load factor of the user-side power consumption. This represents the peak-to-valley difference in power consumption on the user side.
6. The method as described in claim 3, characterized in that, The step of mapping the short-term time-series features and power consumption features to a high-dimensional space for feature fitting and constructing an initial feature function includes: A sample set is constructed by using the features that affect power consumption as input features and the historical power consumption of users as output features. The Gaussian kernel function is selected as the feature mapping function, and the kernel function parameters are tuned on the sample set through cross-validation to obtain the optimal kernel function parameters. A linear regression model is constructed by mapping the input features of the sample set to a high-dimensional space using a Gaussian kernel function. The linear regression model is transformed into its dual form, and the dual problem is solved based on the input and output features of the sample set to obtain the prediction coefficients. The initial feature function is obtained by combining the prediction coefficients with the Gaussian kernel function.
7. The method as described in claim 2, characterized in that, The step of substituting the integrated feature vector into a pre-determined decomposed power prediction model to obtain the power prediction value includes: Discrete wavelet transform is used to decompose the integrated eigenvector into high-frequency abrupt change components and low-frequency periodic components; Gated cyclic units are used to capture the temporal dependence in high-frequency mutation components, and gated linear units are combined to enhance the nonlinear expression capability. A Transformer encoder is used to learn the global dependency of modes in low-frequency periodic components, and time synchronization information is preserved through position encoding; The weights of high-frequency mutation components and low-frequency periodic components are adaptively adjusted based on the mutation intensity at the current moment using learnable weighting coefficients. The power prediction value is obtained from the high-frequency abrupt change component, the low-frequency periodic component, and their respective weights.
8. A short-term power consumption forecasting system for user-side power in a distributed grid-connected system, characterized in that, include: The preprocessing module is used to perform range scaling and standardization on user-side data of distributed photovoltaic grid connection, and to weight and integrate the standardized data to obtain an integrated feature vector. The power prediction module is used to input the integrated feature vector into the pre-built power prediction model to obtain the power prediction value.
9. An electronic device, characterized in that, include: At least one processor and memory; The memory and processor are connected via a bus; The memory is used to store one or more programs; When the one or more programs are executed by the at least one processor, a method for short-term prediction of user-side power consumption in a distributed grid as described in any one of claims 1 to 7 is implemented.
10. A readable storage medium, characterized in that, It contains an executable program, which, when executed, implements a method for short-term prediction of user-side power consumption in a distributed grid as described in any one of claims 1 to 7.