A power distribution network station area load prediction method based on a grey feature mapping model
By using user clustering and frequency domain feature extraction based on a gray feature mapping model, combined with neural network residual correction, the problem that traditional gray prediction methods struggle to characterize user differences and periodic information in transformer area load forecasting is solved. This achieves higher accuracy and more stable load forecasting, improving the reliability and operational efficiency of transformer area load forecasting.
Patent Information
- Application Number
- CN202611299734.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-26
- Publication Date
- 2026-11-10
- Estimated Expiration
- 2046-08-26
AI Technical Summary
Traditional grey forecasting methods are unable to fully characterize the differences in electricity consumption characteristics among different types of users in low-voltage distribution network areas, and they do not make sufficient use of the periodic information contained in the load sequence. This results in limited forecasting accuracy when facing short-term random load fluctuations, affecting the reliability of load uncertainty forecasting in distribution areas.
A method based on a grey feature mapping model is adopted, which combines user clustering, frequency domain feature extraction, grey feature mapping modeling and neural network residual correction to predict the uncertainty of transformer area load by considering the differences, periodicity and random fluctuations of user load.
It improves the accuracy and stability of load forecasting for low-voltage distribution network areas, enhances the adaptability to complex scenarios, and improves the accuracy of load forecasting and the efficiency of operation and scheduling.
Smart Images

Figure CN122823407B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of power system load forecasting technology, and in particular to a method for forecasting distribution network area loads based on a grey feature mapping model. Background Technology
[0002] With the continuous advancement of new power systems and distribution network construction, low-voltage distribution networks, as a crucial link connecting the power system and end users, directly impact power supply reliability, power quality, and grid operational safety. Distribution area load forecasting provides vital data support for the economic operation of the distribution network, demand response, distributed energy consumption, and optimized allocation of distribution equipment, making it a key technology for achieving intelligent dispatching and refined management of the distribution network. Therefore, accurate and efficient uncertainty forecasting of low-voltage distribution network loads has significant theoretical and engineering application value.
[0003] Currently, load forecasting methods for power distribution areas mainly include statistical forecasting methods, machine learning methods, deep learning methods, and grey forecasting methods. Among these, statistical forecasting methods require high data stability and struggle to adapt to the nonlinearity and random fluctuations of load data. While machine learning and deep learning methods possess strong nonlinear fitting capabilities, they typically rely on a large number of high-quality historical samples, resulting in complex model training and high computational costs. Their predictive performance tends to degrade in scenarios with small user load sample sizes, drastic data fluctuations, and incomplete information. Grey forecasting methods, on the other hand, can fully utilize limited sample information and offer advantages such as simple modeling, high computational efficiency, and applicability to small samples and uncertain systems. Therefore, they have been widely used in the field of power load forecasting.
[0004] However, low-voltage distribution network areas typically contain a large number of users, whose electricity consumption behaviors exhibit significant differences and randomness, resulting in complex characteristics of multi-source heterogeneity and coupling of periodicity and randomness in the load of distribution areas. Traditional grey prediction methods usually model the overall load of the distribution area directly, making it difficult to fully characterize the differences in electricity consumption characteristics among different types of users, and they do not make sufficient use of the periodic information contained in the load sequence, resulting in limited prediction accuracy when facing short-term random load fluctuations. In addition, traditional grey models are mainly built on linear or low-dimensional feature spaces, making it difficult to effectively describe complex nonlinear relationships, and model prediction errors are prone to accumulation, affecting the reliability of uncertain load prediction in distribution areas. Therefore, there is an urgent need to propose a new method that can fully explore user load characteristics, enhance the nonlinear representation capability of grey models, and improve the accuracy of uncertain load prediction in low-voltage distribution network areas. Summary of the Invention
[0005] To address the shortcomings of the aforementioned background technologies, this invention proposes a distribution network substation load forecasting method based on a grey feature mapping model. This method fully considers the variability, periodicity, and random fluctuations of user loads. Through user clustering, frequency domain feature extraction, grey feature mapping modeling, neural network residual correction, and substation load aggregation forecasting, it aims to achieve uncertain prediction of short-term loads in low-voltage distribution network substations, improving the accuracy and stability of substation load forecasting in complex scenarios.
[0006] To achieve the above-mentioned technical objectives, the present invention provides the following technical solution:
[0007] A method for load forecasting of distribution network areas based on a grey feature mapping model, which specifically includes the following steps:
[0008] S1. Collect historical load data of all users in the transformer area, perform data cleaning and preprocessing, and construct user load time series;
[0009] S2. Extract user feature vectors from user load time series, and classify all users in the transformer area based on user feature vectors;
[0010] S3. Use Fast Fourier Transform to extract the frequency domain feature vectors of various user load time series;
[0011] S4. Fuse the frequency domain feature vector, time domain load features, and time-series location features into an input feature vector; construct a grey feature mapping model, perform random feature mapping and gated residual learning on the input feature vector, and obtain the initial prediction results of the load of each user in the future period;
[0012] S5. For the prediction residuals generated by the gray feature mapping model, perform iterative training of the model, and perform nonlinear learning and dynamic compensation on the prediction residuals; use the trained gray feature mapping model to obtain the final quantile prediction results of each user's load.
[0013] S6. Perform load aggregation on the final quantile prediction results of each user's load to obtain the uncertainty prediction results of the total load of the low-voltage distribution network area.
[0014] Furthermore, step S1 specifically includes:
[0015] Collect historical load data of users in low-voltage distribution network areas to obtain the original load sequence of users;
[0016] The original user load sequence is processed by linear interpolation to remove outliers and fill in missing values. The data is then standardized using the min-max normalization method. Predictive samples are then constructed using a sliding window method according to a preset time window length to obtain the user load time series.
[0017] Furthermore, step S2 specifically includes:
[0018] Statistical features, time-series features, and electricity consumption behavior features of users' historical load sequences are extracted to construct user feature vectors. The K-means clustering algorithm is used to cluster users based on their feature vectors, dividing all users in the transformer area into different categories.
[0019] Furthermore, step S3 specifically includes:
[0020] The loads of all users in the same category at each sampling time are summed to obtain the aggregated load sequence of that category of users. A fast Fourier transform is performed on the aggregated load sequence of each category of users to obtain the complex spectral coefficients of the aggregated load sequence at each frequency. The amplitude spectrum at each frequency is calculated based on the complex spectral coefficients. The spectral energy corresponding to each frequency component is calculated based on the amplitude spectrum. The top u frequency components with the largest spectral energy are selected as the dominant frequency components of the load time series of each category of users. The frequency domain feature vector of the load time series of each category of users is constructed from the dominant frequency components.
[0021] Further, in step S4, the specific steps of fusing the frequency domain feature vector, time domain load feature, and time-series location feature into an input feature vector are as follows:
[0022] Introducing exponential information to accumulate and generate sequences ,in It is the exponential accumulation coefficient. ; For the current prediction time ; For the first Class 1 The user in the first The electricity load values at each sampling time point are collected; a sequence is generated by accumulating the exponential information, and time-domain load characteristics are extracted, including short-term time-delay characteristics and main periodic characteristics:
[0023] The sequence is generated by accumulating exponential information, and short-term time-lag features are first extracted from the current prediction time backward. The formula is expressed as:
[0024] ;
[0025] in, Indicates the sampling interval. Indicates the first The preset short-term delay length for this type of user;
[0026] By combining the dominant frequency components of various user load time series, the main periodic features are extracted from the current prediction time backward. The formula is expressed as:
[0027] ;
[0028] in, Indicates the first The main periodic sequence of the user load time series is obtained by taking the reciprocal of each main period, corresponding to the main frequency component. Indicates the first The first user load time series There are three main cycles; different short-term time delay feature lengths and main cycle feature lengths are used for different categories of users;
[0029] Then, time encoding is introduced. Based on the intraday sampling point number and weekday number to which the current prediction time belongs, a temporal location feature is constructed, expressed by the formula:
[0030] ;
[0031] in, This indicates that the current prediction time is the p-th sampling point in a day. Indicates the week number;
[0032] By combining short-term time delay features, main periodic features, and temporal position features, the current prediction time can be obtained. Input feature vector .
[0033] Further, in step S4, the construction of the gray feature mapping model, performing random feature mapping and gated residual learning on the input feature vector to obtain the initial prediction results for the load of each user in the future time period, specifically involves:
[0034] A grey feature mapping model GFMM is constructed, which includes a random feature mapping module and a gated residual memory network GRN. The input feature vector is randomized in the random feature mapping module to obtain randomized features. The input feature vector is then linearly mapped, gated fusion, residual connection and layer normalization, feedforward enhancement and output linear mapping in the gated residual memory network to obtain gated residual features.
[0035] The random mapping feature is added to the gated residual feature to obtain the quantile prediction result for the next h steps, which is used as the initial prediction result for user load in the future period.
[0036] Furthermore, step S5 specifically includes:
[0037] The multiquantile joint loss function with L2 regularization is used as the objective function of the grey feature mapping model; for the ... The objective function of the gray feature mapping model is a quantile. for:
[0038] ;
[0039] in, , Indicates the predicted residual; Indicates the first quantiles; The square of the Frobenius norm; This represents the actual load value; Indicates the first Forecast load values at each quantile; These are the L2 regularization coefficients; C is the output matrix of the random mapping.
[0040] The AdamW optimization algorithm is used to optimize the parameters of the gray feature mapping model, and the model parameters are updated iteratively through backpropagation. Using the trained gray feature mapping model, the final quantile prediction results of each user at future time are obtained.
[0041] Furthermore, step S6 specifically includes:
[0042] Within each user category, the median forecast result in the final quantile forecast result of each user is aggregated to obtain the load forecast value of each user category; the load forecast values of each user category are aggregated to obtain the total load forecast value of the low-voltage distribution network area.
[0043] Within each user category, load aggregation is performed on the upper and lower quantile prediction results of the final quantile prediction results for each user to obtain the upper and lower limits of the load prediction interval for each user category; then, the upper and lower limits of the load prediction interval for each user category are aggregated to obtain the upper and lower limits of the prediction interval for the load of the low-voltage distribution network area.
[0044] The uncertainty prediction of the total load of the low-voltage distribution network area includes the total load prediction value and the prediction range.
[0045] Furthermore, this application also discloses a distribution network area load forecasting system based on a grey feature mapping model, which is used to execute the method described above, specifically including:
[0046] The user classification module is used to extract the user feature vectors of the historical load of all residential users in the transformer area, and classify the residential users by using cluster analysis.
[0047] The gray prediction module is used to construct a gray feature mapping model. It obtains the initial prediction results of the user load for future periods through exponential information accumulation, random feature mapping, and gated residual learning.
[0048] The model iterative training module is used to iteratively train the model based on the prediction residuals generated by the grey feature mapping model, and to perform nonlinear learning and dynamic compensation on the prediction residuals; the trained grey feature mapping model is used to obtain the final quantile prediction results of each user's load.
[0049] The load integration module is used to aggregate the final quantile prediction results of various users to obtain the total load prediction value and prediction range of the low-voltage distribution network area, thereby realizing the uncertainty prediction of the area load.
[0050] Based on the above technical solution, the present invention has at least the following beneficial effects:
[0051] 1. This invention uses a user clustering method to classify and model the load of low-voltage distribution network areas, fully considering the differences in electricity consumption behavior of different types of users, and improving the load prediction model's relevance to different types of users and its adaptability to complex load scenarios.
[0052] 2. This invention proposes a grey feature mapping model whose input feature vector includes short-term time-delay features, main periodic features, and time-encoded features. It can characterize load change characteristics from multiple dimensions such as short-term correlation, periodicity, and time attributes. The short-term time-delay features and main periodic features are constructed based on the accumulation of exponential information to generate sequences, which can highlight the role of new information, suppress random fluctuations, improve feature stability, and enhance the regularity of load sequences. At the same time, combined with the fast Fourier transform adaptive extraction of main periodic features and the new information weighted parameter estimation method, the grey feature mapping model's ability to characterize the actual load periodic characteristics and dynamic changes is improved, further enhancing the accuracy of short-term load forecasting.
[0053] 3. This invention uses a clustering aggregation prediction method to aggregate the prediction results of various users, obtain the total load prediction results and prediction range of the low-voltage distribution network area, realize the uncertainty prediction of the area load, and improve the accuracy and efficiency of distribution network operation scheduling and demand response. Attached Figure Description
[0054] Figure 1 This is a flowchart illustrating a method for predicting load in distribution network areas based on a grey feature mapping model proposed in this invention.
[0055] Figure 2 This is a comparison chart of the statistical characteristic curves of the load of different categories of users in the embodiment;
[0056] Figure 3 This is a comparison chart of the accuracy of the load prediction experiment for the substation area in the embodiment;
[0057] Figure 4 This is a box plot comparing the prediction errors of the present invention and existing models. Detailed Implementation
[0058] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, but the scope of protection of the present invention is not limited to the following embodiments.
[0059] It should be understood that any modifications, equivalent substitutions, and improvements made to this invention by those skilled in the art without departing from the spirit and scope of this invention shall fall within the scope of protection of this invention.
[0060] Although the steps in this invention are arranged by reference numerals, this is not intended to limit the order of the steps. Unless the order of the steps is explicitly stated or the execution of a step requires other steps as a basis, the relative order of the steps can be adjusted. It is understood that the term "and / or" as used herein refers to and covers any and all possible combinations of one or more of the associated listed items.
[0061] like Figure 1 As shown, this invention proposes a load forecasting method for distribution network areas based on a grey feature mapping model, which specifically includes the following steps:
[0062] S1. Collect historical load data of all users in the transformer area, perform data cleaning and preprocessing, and construct user load time series;
[0063] In a preferred embodiment, step S1 specifically comprises:
[0064] Historical load data of users in a low-voltage distribution network area are collected to obtain the original load sequence of users; in this embodiment, actual data of electricity meters of 267 users in a certain area are collected.
[0065] Data quality screening of the original user load sequence is performed using linear interpolation, including:
[0066] Outlier removal was performed, removing users whose historical load data were all zero and users whose missing data ratio exceeded a preset threshold of 20%, ultimately retaining 213 valid user data; for the small number of missing data among the retained users, linear interpolation was used to fill in the missing values;
[0067] Then, the data is standardized using the min-max normalization method, and then a prediction sample is constructed using a sliding window method according to the preset time window length to obtain the user load time series.
[0068] ;
[0069] in, This represents the load time series of the i-th user. Indicates the i-th user's... The load value at each sampling time. This represents the total number of historical sampling times. In this embodiment, the historical sampling time window length is set to 2025.06--2025.09, and the sampling frequency is 15 minutes.
[0070] S2. Extract user feature vectors from user load time series, and classify all users in the transformer area based on user feature vectors;
[0071] In a preferred embodiment, step S2 specifically comprises:
[0072] Extract the statistical features, time-series features, and electricity consumption behavior features of users' historical load sequences to construct user feature vectors, expressed by the following formula:
[0073] ;
[0074] Where m represents the feature dimension;
[0075] The K-means clustering algorithm is used to cluster users, dividing all users within the station area into different categories; the clustering objective function is... Represented as:
[0076] ;
[0077] Where K is the number of clusters, Let j represent the set of users of the jth class. This indicates the corresponding cluster center, thus obtaining different types of users.
[0078] After clustering, the 213 valid user data in this area can be roughly divided into four categories. As shown in Table 1, there are 130 residential users, accounting for 61.03%; 46 charging pile users, accounting for 21.60%; 31 photovoltaic users, accounting for 14.55%; and 6 industrial and commercial users, accounting for 2.82%.
[0079] Table 1 User Classification Results
[0080]
[0081] Figure 2 The chart shows a comparison of the daily load statistical characteristic curves for the four types of users mentioned above. Figure 2 (a) in the figure is the daily load statistical characteristic curve of photovoltaic users. The load curve has a large number of negative values, showing obvious diurnal periodicity. This is because the grid connection of photovoltaic power generation during the day leads to the reverse flow of electricity, while at night it is mainly the user's own electricity consumption. Therefore, the load changes between positive and negative values, which has typical photovoltaic output characteristics. Figure 2(b) in the figure represents the daily load statistical characteristic curve of residential users. The load is mainly positive and has strong random fluctuations. The peak value mainly comes from the electricity consumption of residents in daily life. Overall, it shows the characteristics of low base load, random fluctuations and intermittent peak values. Figure 2 (c) in the figure represents the daily load statistical characteristic curve of industrial and commercial users. The load curve is relatively stable overall with small fluctuations, exhibiting high continuity and stability. It only fluctuates slightly during certain periods due to production activities, reflecting the strong regularity of electricity consumption characteristics of industrial and commercial users. Figure 2 (d) in the table represents charging pile users, whose load exhibits typical intermittent pulse characteristics. Most of the time, they are in a zero-load state, and only when the vehicle is charging does a sustained high load occur. Therefore, they have the characteristics of sparseness, high peak value and random start, which is significantly different from other types of users.
[0082] It can be seen that there are significant differences in the electricity consumption behavior of different types of users. Therefore, this application models the input feature vectors of user load time series in the gray feature mapping model according to categories, so as to improve the load prediction model’s targeting of different types of users and its adaptability to complex load scenarios.
[0083] S3. Use Fast Fourier Transform to extract the frequency domain feature vectors of various user load time series;
[0084] In a preferred embodiment, step S3 specifically comprises:
[0085] The load of all users in the same category at each sampling time is summed to obtain the aggregated load sequence of that category of users; a Fast Fourier Transform is performed on the aggregated load sequence of each category of users to obtain the complex spectral coefficients of the aggregated load sequence at each frequency; for the q-th category of users, its aggregated load sequence is expressed as... Its fast Fourier transform is expressed as:
[0086] ;
[0087] in, Represents frequency Complex spectral coefficients at [location] This represents the sequence length, where z is the imaginary unit. For user of type q in the th case Aggregated load sequence at each sampling time;
[0088] Calculate the amplitude spectrum at each frequency based on complex spectral coefficients. The spectral energy corresponding to each frequency component is calculated based on the amplitude spectrum, and the top u frequency components with the largest spectral energy are selected as the dominant frequency components of the time series of various user loads. The set of dominant frequency components is represented as follows: Frequency domain feature vectors of various user load time series are constructed from the main frequency components.
[0089] Furthermore, as shown in Table 2, statistical analysis of the main frequency components of the load time series of various types of users revealed that different types of users all exhibit obvious periodic electricity consumption patterns.
[0090] Table 2 Statistical Analysis Results of User Data Frequency
[0091]
[0092] Among them, the 24-hour cycle occurred most frequently, accounting for 98.12% of all users, indicating that the vast majority of users have stable daily cycle characteristics; the 12-hour, 8-hour, and 6-hour cycles accounted for 69.48%, 53.52%, and 28.64% respectively, indicating that user load has both semi-daily cycle and multi-scale short-cycle fluctuation characteristics. Therefore, when designing the input feature vector of the gray feature mapping model, this periodic electricity consumption pattern and multi-scale short-cycle fluctuation should be taken into account.
[0093] S4. Fuse the frequency domain feature vector, time domain load features, and time-series location features into an input feature vector; construct a grey feature mapping model, perform random feature mapping and gated residual learning on the input feature vector, and obtain the initial prediction results of the load of each user in the future period;
[0094] In a preferred embodiment, step S4 specifically comprises:
[0095] To mitigate random fluctuations in the original load sequence, exponential information is introduced to accumulate and generate the sequence. ,in It is the exponential accumulation coefficient. The default value is 0.5; For the current prediction time ; For the first Class 1 The user in the first The electricity load values at each sampling time point are collected; a sequence is generated by accumulating the exponential information, and time-domain load characteristics are extracted, including short-term time-delay characteristics and main periodic characteristics:
[0096] The sequence is generated by accumulating exponential information, and short-term time-lag features are first extracted from the current prediction time backward. The formula is expressed as:
[0097] ;
[0098] in, Indicates the sampling interval. Indicates the first The preset short-term delay length for this type of user;
[0099] By combining the dominant frequency components of various user load time series, the main periodic features are extracted from the current prediction time backward. The formula is expressed as:
[0100] ;
[0101] in, Indicates the first The main periodic sequence of the user load time series is obtained by taking the reciprocal of each main period, corresponding to the main frequency component. Indicates the first The first user load time series There are four main periods; different user categories use different short-term time delay feature lengths and main period feature lengths. In this embodiment, taking user type as an example, a short-term time delay feature with a length of 3 hours is taken, and a main period feature with four main period lengths is taken, that is:
[0102] ;
[0103] ;
[0104] To characterize the intraday and weekly cycles of load, time coding is introduced. Based on the intraday sampling point number and weekday number of the current prediction time, a time-series location feature is constructed, expressed by the formula:
[0105] ;
[0106] in, This indicates that the current prediction time is the p-th sampling point in a day. Indicates the week number;
[0107] Since the sampling frequency in this embodiment is 15 minutes, therefore:
[0108] ;
[0109] By combining short-term time delay features, main periodic features, and temporal position features, the current prediction time can be obtained. Input feature vector .
[0110] Next, a grey feature mapping model GFMM is constructed, which includes a stochastic feature mapping module and a gated residual memory network GRN. The stochastic feature mapping module is used to enhance the grey model's ability to express complex nonlinear load changes, and the gated residual learning module is used to learn the load change patterns that are difficult to express by stochastic mapping.
[0111] The input feature vector undergoes random feature mapping in the random feature mapping module to obtain randomly mapped features, expressed by the formula:
[0112] ;
[0113] in, It is the mapping function of the random feature mapping algorithm. That is to indicate input The resulting random mapping features; It is a random mapping matrix. For random bias vectors, The dimension of the random feature is... Output matrix for random mapping;
[0114] The input feature vector undergoes a linear mapping process in the gated residual memory network, as expressed by the formula:
[0115] ;
[0116] in, express Output features via linear mapping;
[0117] Gating fusion is performed, and the formula is expressed as:
[0118] ;
[0119] in, For gated output, gate vector State vector , Represents the Hadamard product
[0120] Then, through residual connections and layer normalization, the hidden state is obtained. The formula is expressed as:
[0121] ;
[0122] Then, feedforward enhancement is applied to the hidden state, as expressed by the formula:
[0123] ;
[0124] in, This is a feedforward enhancement feature;
[0125] Finally, the feedforward enhanced features are linearly mapped to the output to obtain the gated residual features. Its output dimension is consistent with that of the random mapping branch, as expressed by the formula:
[0126] ;
[0127] in, , , , and The weight matrix for each layer of the network. , , , and These are the bias vectors for each layer; This represents a feedforward neural network; Presentation layer normalization operation; This represents the Sigmoid activation function; This represents the hyperbolic tangent activation function.
[0128] Adding the random mapping features to the gated residual features yields the quantile prediction result for the next h steps. The formula is expressed as:
[0129] ;
[0130] This serves as an initial prediction of user load for future periods. The difference between different quantile predictions lies in the different values of the parameters in the grey feature mapping model. Subsequently, in step S5, by setting different objective functions for different quantiles, different model parameters can be obtained, leading to different quantile prediction results.
[0131] S5. For the prediction residuals generated by the gray feature mapping model, perform iterative training of the model, and perform nonlinear learning and dynamic compensation on the prediction residuals; use the trained gray feature mapping model to obtain the final quantile prediction results of each user's load.
[0132] In a preferred embodiment, step S5 specifically comprises:
[0133] The multiquantile joint loss function with L2 regularization is used as the objective function of the grey feature mapping model; for the ... The objective function of the gray feature mapping model is a quantile. for:
[0134] ;
[0135] in, , Indicates the predicted residual; Indicates the first quantiles; The square of the Frobenius norm; This represents the actual load value; Indicates the first Forecast load values at each quantile; These are the L2 regularization coefficients; C is the output matrix of the random mapping.
[0136] The AdamW optimization algorithm is used to optimize the parameters of the gray feature mapping model, and the model parameters are updated iteratively through backpropagation. Using the trained gray feature mapping model, the final quantile prediction results of each user at future time are obtained.
[0137] S6. Perform load aggregation on the final quantile prediction results of each user's load to obtain the uncertainty prediction results of the total load of the low-voltage distribution network area;
[0138] In a preferred embodiment, step S6 specifically comprises:
[0139] Based on the user clustering results, users within the station area are divided into: Class of users, number The number of users of each class is denoted as Then there is ,in, This indicates the total number of active users within the designated area.
[0140] Within each user category, the median prediction result (i.e., using the median prediction result as the load prediction value for each user, in this embodiment, the 0.5 quantile is taken as the median) from the final quantile prediction results for each user is aggregated to obtain the load prediction values for each user category; let the quantile be denoted as _____. Class 1 The predicted load value for each user at prediction time t is Then the first The load forecast value for user class is expressed as ;
[0141] Then, the load forecasts for various users are aggregated to obtain the total load forecast for the low-voltage distribution network area. .
[0142] Within each user category, load aggregation is performed on the upper and lower quantile prediction results of each user's final quantile prediction results to obtain the upper and lower limits of the load prediction interval for each user category. Then, the upper and lower limits of the load prediction interval for each user category are aggregated to obtain the upper and lower limits of the prediction interval for the load of the low-voltage distribution network area. The formula is expressed as:
[0143] ;
[0144] ;
[0145] in, For the first Class 1 The lower quantile prediction results for each user at prediction time t. For the first Class 1 The upper quantile prediction result for each user at prediction time t. In this embodiment, the lower quantile and upper quantile are set to 0.1 and 0.9, respectively. , These are the lower and upper limits of the predicted load range for the low-voltage distribution area, respectively.
[0146] This enables the prediction of load uncertainty in distribution areas, and the final prediction result of load uncertainty in low-voltage distribution network distribution areas is expressed as follows: .
[0147] This concludes the description of the entire process of the method proposed in this invention.
[0148] Furthermore, this embodiment also uses 130 residential households as an example to compare the experimental results of the method proposed in this invention with other classic prediction methods; such as Figure 3 The chart shows a comparison of the accuracy of different models in load forecasting experiments for residential users. Figure 3 (a) Figure 3 (b) Figure 3 (c) Figure 3 (d) Figure 3 (e) in Figure 3 (f) shows line graphs comparing the predicted load values of residential users with the actual load values for the proposed Grey Feature Mapping Model (GFMM), Recurrent Neural Network (RNN), Long Short-Term Memory Network (LSTM), Transformer, Frequency Enhanced Decomposed Transformer (FEDformer), and Autoregressive Integrated Moving Average (ARIMA) models.
[0149] The average prediction error of different models is calculated based on the predicted and actual values. As shown in Table 3 below, the average prediction error of different models is compared. Among them, RMSE (Root Mean Square Error), MAE (Mean Absolute Error), and MAPE (Mean Absolute Percentage Error) are used to evaluate the prediction accuracy of the model. The smaller the index value, the better the prediction performance of the model.
[0150] Table 3. Average Error of Comparison Models
[0151]
[0152] As shown in Table 3, the GFMM model achieved the best results in all three evaluation metrics: RMSE, MAE, and MAPE. The average RMSE, MAE, and MAPE were 0.3591, 0.1699, and 78.6917%, respectively, all outperforming other comparative models. Specifically, compared to the best-performing deep learning model, Transformer, GFMM reduced MAE and MAPE by 16.8% and 37.9%, respectively; and compared to the traditional ARIMA model, reduced them by 50.9% and 71.8%, respectively. These results indicate that GFMM can more accurately characterize the nonlinear and time-varying features of residential load sequences, demonstrating a significant advantage in prediction accuracy.
[0153] Figure 4 This shows the box plot of the prediction error between the present invention and existing models. From Figure 4 As can be seen, all models can reflect the load change trend, but there are significant differences in prediction accuracy and stability. Among them, the GFMM model proposed in this invention has the highest degree of agreement between its prediction curve and the actual value, and can accurately track the peak and trough changes of the load. Existing models such as RNN, LSTM, Transformer, and FEDformer can learn the temporal characteristics of the load, but still have certain deviations in local peaks and rapid fluctuations. The ARIMA model has a relatively large prediction error and is insufficient in characterizing complex nonlinear load changes. At the same time, the GFMM model has the lowest median MAE and the most concentrated error distribution, showing higher prediction accuracy and better stability.
[0154] In summary, this invention addresses the problems of strong random fluctuations, complex periodic characteristics, and insufficient prediction accuracy of traditional forecasting methods under small sample conditions in low-voltage distribution network load forecasting. It proposes a load forecasting method for distribution network areas based on a grey feature mapping model. This method fully utilizes the advantages of grey system theory in small sample modeling, combining user clustering, periodic feature extraction, random feature mapping, and residual optimization techniques to achieve in-depth mining and nonlinear modeling of user load characteristics. Through clustering and aggregation forecasting, it obtains the total load forecast results and forecast intervals for the distribution area, forming a complete technical solution suitable for uncertain load forecasting in low-voltage distribution networks. This method can provide reliable decision-making basis for distribution network operation scheduling, demand response, and coordinated operation of distributed energy resources, and has good engineering application value.
[0155] In this specification, the terms "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to at least one embodiment or example described in conjunction with specific features, structures, steps, or characteristics. The specific features, structures, steps, or characteristics described in this specification can be combined in any suitable manner among multiple embodiments or examples. Without contradiction, those skilled in the art can combine, substitute, or modify the different embodiments or examples disclosed in this specification.
[0156] The method flow involved in this invention can be implemented by software, hardware, or a combination of both. Each step shown in the flowchart or otherwise described can be understood as a series of program instructions executable by a processor. These program instructions can be stored in a computer-readable storage medium and loaded and executed by the processor to complete the low-voltage distribution network area load uncertainty prediction method described in this invention. The computer-readable storage medium includes, but is not limited to, read-only memory (ROM), random access memory (RAM), magnetic disk, optical disk, flash memory, or other storage media capable of storing program code.
[0157] The above embodiments have provided a detailed description of the technical solution of the present invention, and the principles of the present invention have been explained in conjunction with specific examples. The purpose of this description is to help understand the core ideas and technical solutions of the present invention. Those skilled in the art can make various modifications, equivalent substitutions, or improvements to the technical solution of the present invention without departing from the principles and spirit of the present invention. All equivalent changes made using the technical concept of the present invention should fall within the protection scope defined by the claims of the present invention.
Claims
1. A method for load forecasting of distribution network areas based on a grey feature mapping model, characterized in that, Specifically, the following steps are included: S1. Collect historical load data of all users in the transformer area, perform data cleaning and preprocessing, and construct user load time series; S2. Extract user feature vectors from user load time series, and classify all users in the transformer area based on user feature vectors; S3. Use Fast Fourier Transform to extract the frequency domain feature vectors of various user load time series; S4. Fuse the frequency domain feature vector, time domain load features, and time-series location features into an input feature vector; construct a grey feature mapping model, perform random feature mapping and gated residual learning on the input feature vector, and obtain the initial prediction results for the load of each user in the future period; specifically: A grey feature mapping model GFMM is constructed, which includes a random feature mapping module and a gated residual memory network GRN. The input feature vector is randomized in the random feature mapping module to obtain randomized features. The input feature vector is then linearly mapped, gated fusion, residual connection and layer normalization, feedforward enhancement and output linear mapping in the gated residual memory network to obtain gated residual features. The random mapping feature is added to the gated residual feature to obtain the quantile prediction result for the next h steps, which is used as the initial prediction result for user load in the future period. S5. For the prediction residuals generated by the gray feature mapping model, perform iterative training of the model, and perform nonlinear learning and dynamic compensation on the prediction residuals; use the trained gray feature mapping model to obtain the final quantile prediction results of each user's load. S6. Perform load aggregation on the final quantile prediction results of each user's load to obtain the uncertainty prediction result of the total load of the low-voltage distribution network area; Step S6 specifically includes: Within each user category, the median forecast result in the final quantile forecast result of each user is aggregated to obtain the load forecast value of each user category; the load forecast values of each user category are aggregated to obtain the total load forecast value of the low-voltage distribution network area. Within each user category, load aggregation is performed on the upper quantile and lower quantile prediction results of each user's final quantile prediction results to obtain the upper and lower limits of the load prediction interval for each user category. Then, the upper and lower limits of the load forecast intervals for various types of users are aggregated to obtain the upper and lower limits of the load forecast intervals for low-voltage distribution network areas; The uncertainty prediction of the total load of the low-voltage distribution network area includes the total load prediction value and the prediction range.
2. The method for load forecasting of distribution network areas based on a grey feature mapping model according to claim 1, characterized in that, Step S1 is as follows: Collect historical load data of users in low-voltage distribution network areas to obtain the original load sequence of users; The original user load sequence is processed by linear interpolation to remove outliers and fill in missing values. The data is then standardized using the min-max normalization method. Predictive samples are then constructed using a sliding window method according to a preset time window length to obtain the user load time series.
3. The method for load forecasting of distribution network areas based on a grey feature mapping model according to claim 1, characterized in that, Step S2 is as follows: Statistical features, time-series features, and electricity consumption behavior features of users' historical load sequences are extracted to construct user feature vectors. The K-means clustering algorithm is used to cluster users based on their feature vectors, dividing all users in the transformer area into different categories.
4. The method for load forecasting of distribution network areas based on a grey feature mapping model according to claim 1, characterized in that, Step S3 is as follows: The loads of all users in the same category at each sampling time are summed to obtain the aggregated load sequence of that category of users. A fast Fourier transform is performed on the aggregated load sequence of each category of users to obtain the complex spectral coefficients of the aggregated load sequence at each frequency. The amplitude spectrum at each frequency is calculated based on the complex spectral coefficients. The spectral energy corresponding to each frequency component is calculated based on the amplitude spectrum. The top u frequency components with the largest spectral energy are selected as the dominant frequency components of the load time series of each category of users. The frequency domain feature vector of the load time series of each category of users is constructed from the dominant frequency components.
5. The method for load forecasting of distribution network areas based on a grey feature mapping model according to claim 4, characterized in that, In step S4, the specific steps of fusing the frequency domain feature vector, time domain load feature, and time-series location feature into an input feature vector are as follows: Introducing exponential information to accumulate and generate sequences ,in It is the exponential accumulation coefficient. ; For the current prediction time ; For the first Class 1 The user in the first The electricity load values at each sampling time point are collected; a sequence is generated by accumulating the exponential information, and time-domain load characteristics are extracted, including short-term time-delay characteristics and main periodic characteristics: First, extract short-term time-delay features from the current prediction time backward. The formula is expressed as: ; in, Indicates the sampling interval. Indicates the first The preset short-term delay length for this type of user; By combining the dominant frequency components of various user load time series, the main periodic features are extracted from the current prediction time backward. The formula is expressed as: ; in, Indicates the first The main periodic sequence of the user load time series is obtained by taking the reciprocal of each main period, corresponding to the main frequency component. Indicates the first The first user load time series There are three main cycles; different short-term time delay feature lengths and main cycle feature lengths are used for different categories of users; Then, time encoding is introduced. Based on the intraday sampling point number and weekday number to which the current prediction time belongs, a temporal location feature is constructed, expressed by the formula: ; in, This indicates that the current prediction time is the p-th sampling point in a day. Indicates the week number; By combining short-term time delay features, main periodic features, and temporal position features, the current prediction time can be obtained. Input feature vector .
6. The method for load forecasting of distribution network areas based on a grey feature mapping model according to claim 1, characterized in that, Step S5 is as follows: The multiquantile joint loss function with L2 regularization is used as the objective function of the grey feature mapping model; for the ... The objective function of the gray feature mapping model is: quantiles. for: ; in, , Indicates the predicted residual; Indicates the first quantiles; The square of the Frobenius norm; This represents the actual load value; Indicates the first Forecast load values at each quantile; These are the L2 regularization coefficients; C is the output matrix of the random mapping. The AdamW optimization algorithm is used to optimize the parameters of the gray feature mapping model, and the model parameters are updated iteratively through backpropagation. Using the trained gray feature mapping model, the final quantile prediction results of each user at future time are obtained.
7. A distribution network substation load forecasting system based on a grey feature mapping model, the system being used to execute the method as described in any one of claims 1-6, characterized in that, Specifically, it includes: The user classification module is used to extract the user feature vectors of the historical load of all residential users in the transformer area, and classify the residential users by using cluster analysis. The gray prediction module is used to construct a gray feature mapping model. It obtains the initial prediction results of the user load for future periods through exponential information accumulation, random feature mapping, and gated residual learning. The model iterative training module is used to iteratively train the model based on the prediction residuals generated by the grey feature mapping model, and to perform nonlinear learning and dynamic compensation on the prediction residuals; the trained grey feature mapping model is used to obtain the final quantile prediction results of each user's load. The load integration module is used to aggregate the final quantile prediction results of various users to obtain the total load prediction value and prediction range of the low-voltage distribution network area, thereby realizing the uncertainty prediction of the area load.
Citation Information
Patent Citations
Transformer area load regulation and control prediction method based on multi-gating hybrid long short-term memory network
CN119382065A
Urbanized power distribution network ultra-short-term load prediction method based on deep embedded clustering and ICEEMDAN-iTransform
CN122092212A
