Short-term load prediction method for power distribution network
By employing core data classification, time-series decomposition, and feature fusion methods, combined with GCN and Transformer models, the problem of load forecasting deviation caused by data gaps and reverse power flow in distribution networks was solved, achieving highly fault-tolerant short-term load forecasting.
Patent Information
- Application Number
- CN202511691897.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-18
- Publication Date
- 2026-03-17
AI Technical Summary
Existing short-term load forecasting methods rely too heavily on data quality in distribution networks, resulting in poor fault tolerance. Load data is lost when end-user meters fail or communication links are interrupted. Furthermore, the reverse power flow caused by distributed photovoltaic and energy storage devices leads to discrepancies between traditional load data and actual electricity demand. Existing methods cannot effectively restore data authenticity and retain key fluctuation characteristics.
We employ a method that combines core data classification and collection, Z-score standardization, temporal decomposition, feature extraction and fusion, and GCN and Transformer models. By calculating feature correlation through mutual information entropy and dynamically adjusting weights, we combine a combined loss function and manually constructed missing data samples for model training, thereby achieving accurate repair and prediction of missing data and reverse flow.
In scenarios such as meter failure and communication interruption, the accuracy and fault tolerance of load forecasting have been improved, ensuring the accuracy and reliability of forecast results and reducing forecast deviation.
Smart Images
Figure CN121688822A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of power system load forecasting technology, specifically a short-term load forecasting method for distribution networks. Background Technology
[0002] With the continuous development of grid-connected photovoltaic (PV) power generation technology and increased policy support, the cost of PV power generation has been greatly reduced, and the penetration rate of distributed PV power generation in distribution networks is constantly increasing. This has led to significant changes in the power load of distribution networks, impacting network planning and power dispatch. At the same time, the intermittent and random nature of distributed PV makes load forecasting more complex. If the impact of distributed PV changes on the power system load is not considered, and load forecasting is based solely on grid-supplied load, significant errors will inevitably occur. Therefore, accurate forecasting after grid connection of distributed PV is essential. Consequently, current PV power generation technologies require the use of short-term load forecasting methods.
[0003] However, existing short-term load forecasting methods still have the following technical problems when used:
[0004] Currently, load forecasting relies too heavily on data quality, resulting in poor fault tolerance. In distribution network scenarios, frequent end-user meter failures and communication link interruptions can easily lead to missing load data. Simultaneously, the integration of distributed photovoltaic and energy storage devices generates reverse power flows, causing discrepancies between traditional load data and actual electricity demand. Existing methods often employ simple interpolation to fill in the data or directly remove outliers, which neither restores the data's accuracy nor preserves key fluctuation characteristics, ultimately leading to a systematic shift in forecast results.
[0005] Therefore, a short-term load forecasting method for distribution networks is proposed to address the aforementioned problems. Summary of the Invention
[0006] The purpose of this invention is to provide a short-term load forecasting method for power distribution networks to solve the problems mentioned in the background art.
[0007] To achieve the above objectives, the present invention provides the following technical solution: a short-term load forecasting method for distribution networks, the specific steps of which are as follows:
[0008] Step 1: Core Data Classification and Collection: Classify and collect load-related data and influencing factor data;
[0009] Step 2: Data standardization: Z-score standardization is used to unify the scale of data with different dimensions;
[0010] Step 3: Load data component decomposition: The load data is split into three parts using a time-series decomposition algorithm;
[0011] Step 4: Precise Repair of Each Component: Determine the type of data problem, such as short-term missing data, long-term missing data, reverse flow anomaly, or random outliers.
[0012] Step 5: Feature Extraction Dimensions: Extract feature dimensions for user load characteristics, power supply impact characteristics, and correlation characteristics;
[0013] Step 6, Feature Selection and Fusion: The correlation between features and net load is calculated using mutual information entropy. Features with a correlation ≥ 0.6 are retained. Multi-source features are integrated using a weighted fusion formula, with the weights dynamically adjusted based on historical prediction errors.
[0014] Step 7, Model Structure Design: The model structure is designed as an input layer, an encoding layer, a decoding layer, and an output layer; the input layer receives the fused feature sequence; the encoding layer uses GCN to model the load correlation between stations and extracts spatial correlation features; the decoding layer uses the Transformer multi-head self-attention mechanism to capture temporal dependencies; and the output layer is a fully connected layer that outputs the 15-minute load forecast values for the next 24 / 48 / 72 hours.
[0015] Step 8: Model Training and Optimization: Divide the historical data into training, validation, and test sets in a 7:2:1 ratio. Add manually constructed missing data samples to enhance robustness. Use a combined loss function and perform iterative training with an initial learning rate of 0.001, which decays by 10% every 100 rounds. Training stops when the validation set error does not decrease for 5 consecutive rounds.
[0016] Step 9, Multi-period rolling forecast: At 6:00 AM every day, the load from 7:00 AM to 7:00 AM the next day is predicted based on the latest 24-hour data; it is updated every 4 hours, and the subsequent forecast results are adjusted by combining real-time load and meteorological correction values. When the deviation between real-time load and forecast value is ≥10%, adaptive correction is triggered.
[0017] Step 10, Forecast Result Output: The forecast results are output as basic and auxiliary. The basic output is the load forecast value for the next 1-7 days every 15 minutes and the forecast error range. The auxiliary output is the period when the load peak occurs and the distributed power source absorption suggestion. During the result output process, when the forecast load exceeds 80% of the line carrying capacity, an overload warning is triggered and a load transfer suggestion is pushed.
[0018] Preferably, in step one, the load-related data includes historical load data and real-time load data of the target transformer area / region collected through the power distribution terminal, and the timestamps and fault types of meter failures and communication interruptions are recorded.
[0019] The influencing factors include meteorological data, user characteristic data, and power grid status data.
[0020] Preferably, in step two, the standardized value is calculated using the following formula: Standardized value = (Original value - Mean) / Standard deviation; where the mean and standard deviation are calculated based on three years of historical data.
[0021] Preferably, in step three, the time-series decomposition algorithm is: Load(Xi) = Trend(Xi) + Seasonal(Xi) + Residual(Xi);
[0022] Trend component: reflects the long-term growth / decline trend of load; Seasonal component: reflects intraday and weekly periodicity; Residual component: the fluctuation part affected by sudden factors.
[0023] Preferably, in step five, the user load characteristics include the daily load peak-valley difference, weekly average load fluctuation rate, industry electricity consumption ratio, and holiday load attenuation rate; the power supply impact characteristics include the photovoltaic output forecast value, energy storage charging and discharging status, and the probability of reverse power flow; the correlation characteristics include using fast Fourier transform to convert the time-domain load into the frequency domain and extracting high-frequency correlation components with temperature and photovoltaic output.
[0024] Preferably, in step six, the weighted fusion formula is: Fused_Feature = 0.4 × load characteristic + 0.3 × power supply characteristic + 0.2 × meteorological characteristic + 0.1 × user characteristic.
[0025] Preferably, in step eight, the combined loss function is Loss = MSE + 0.1 × MAE.
[0026] Compared with the prior art, the beneficial effects of the present invention are:
[0027] This invention addresses the core issues of traditional methods—high reliance on data quality, poor fault tolerance, and insufficient bias correction—through a three-tiered logic: component repair to resolve data gaps, feature decoupling to correct reverse power flow deviations, and sample enhancement and dynamic correction to improve fault tolerance. The entire process revolves around restoring true load characteristics, forming a closed-loop anti-interference design from data acquisition to prediction output, ensuring accurate load prediction results even in scenarios involving meter failures, communication interruptions, and frequent reverse power flow. Attached Figure Description
[0028] Figure 1 This is a flowchart of the steps in this prediction method. Detailed Implementation
[0029] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0030] Example:
[0031] Please see Figure 1 The present invention provides a technical solution:
[0032] A short-term load forecasting method for distribution networks, the specific steps of which are as follows:
[0033] Step 1: Core Data Classification and Collection: Classify and collect load-related data and influencing factor data;
[0034] Step 2: Data standardization: Z-score standardization is used to unify the scale of data with different dimensions;
[0035] Step 3: Load data component decomposition: The load data is split into three parts using a time-series decomposition algorithm; the load is then decomposed into trend, seasonal, and residual components.
[0036] Step 4: Precise Repair of Each Component: Determine the type of data problem, identifying whether it is short-term missing data, long-term missing data, reverse flow anomaly, or random outlier. For short-term missing data, the trend component is interpolated using trend data from similar transformer areas, and the seasonal component reuses historical data from the same period. Unlike simple linear interpolation that smooths out fluctuations, this method repairs each component separately. The seasonal component directly reuses periodic features, the trend component references similar transformer areas to avoid bias, and the residual component is smoothed using moving averages. This restores the periodic fluctuations of the load while preserving reasonable fluctuation details, avoiding feature loss caused by missing data.
[0037] The repair method for short-term missing data is as follows:
[0038] Trend component: Interpolated using trend data from similar transformer substations;
[0039] Seasonal component: Reuse historical data from the same period;
[0040] The long-term missing repair method is: to generate trend completion values based on the GCN model by fusing meteorological and user data, and to superimpose seasonal components with period correction coefficients;
[0041] The method to repair reverse power flow anomalies is as follows: separate the actual user load from the output of distributed power sources, where the actual load = metered load + reverse power flow value, and then reconstruct the components;
[0042] The method for repairing random outliers is to identify outliers using the 3σ criterion and replace outliers in the residual components with moving averages.
[0043] Step four addresses long-term missing data by fusing meteorological and user characteristic data using the GCN model to generate trend completion values. At the same time, a periodic correction coefficient is superimposed. Long-term missing data cannot be repaired by a single interpolation. The GCN model can explore the potential correlation between meteorology, user behavior, and load to generate trend values that conform to physical rules. Combined with the periodicity of seasonal components, it can achieve accurate completion of long-term missing data and avoid feature loss caused by directly removing outliers.
[0044] Step 5: Feature Extraction Dimensions: Extract feature dimensions for user load characteristics, power supply impact characteristics, and correlation characteristics;
[0045] Step 6, Feature Selection and Fusion: The correlation between features and net load is calculated using mutual information entropy. Features with a correlation ≥ 0.6 are retained. Multi-source features are integrated using a weighted fusion formula, with the weights dynamically adjusted based on historical prediction errors.
[0046] Step 6: Select features with a correlation of ≥0.6 with net load through mutual information entropy, and integrate them using a weighted fusion formula. The impact of reverse power flow has temporal differences. Dynamic weight design allows the model to focus on the corresponding features at different times, accurately quantifying the interference of power output on load data.
[0047] Step 7, Model Structure Design: The model structure is designed as an input layer, encoding layer, decoding layer, and output layer. The input layer receives the fused feature sequence with a time step of 96, corresponding to 15-minute granular data per day. The encoding layer models the load correlation between stations using GCN, constructs a station topology map, updates the adjacency matrix based on line impedance, and extracts spatial correlation features. The decoding layer uses the Transformer multi-head self-attention mechanism to capture temporal dependencies. The fully connected output layer outputs the 15-minute load forecast values for the next 24 / 48 / 72 hours.
[0048] Step 8: Model Training and Optimization: Divide the historical data into training, validation, and test sets in a 7:2:1 ratio. Add manually constructed missing data samples to enhance robustness. Use a combined loss function and perform iterative training with an initial learning rate of 0.001, which decays by 10% every 100 rounds. Training stops when the validation set error does not decrease for 5 consecutive rounds.
[0049] Step eight involves adding manually constructed missing data samples during model training. This allows the model to learn the logic for repairing missing data during the training phase. Traditional models, which are not trained on missing data, are prone to inaccuracy in real-world scenarios. By creating artificial samples, the model can adapt to abnormal scenarios such as missing data and reverse currents in advance, improving the fault tolerance of predictions and avoiding prediction collapse due to fluctuations in data quality.
[0050] Step 9, Multi-period rolling forecast: At 6:00 AM every day, the load from 7:00 AM to 7:00 AM the next day is predicted based on the latest 24-hour data; it is updated every 4 hours, and the subsequent forecast results are adjusted by combining real-time load and meteorological correction values. When the deviation between real-time load and forecast value is ≥10%, adaptive correction is triggered.
[0051] Step 10, Forecast Result Output: The forecast results are output as basic and auxiliary. The basic output is the load forecast value for the next 1-7 days every 15 minutes and the forecast error range. The auxiliary output is the period when the load peak occurs and the distributed power source absorption suggestion. During the result output process, when the forecast load exceeds 80% of the line carrying capacity, an overload warning is triggered and a load transfer suggestion is pushed.
[0052] In step one, the load-related data includes historical load data and real-time load data of the target transformer area / region collected through the power distribution terminal, and the timestamps and fault types of meter failures and communication interruptions are recorded.
[0053] Influencing factors include meteorological data (e.g., hourly temperature, humidity, and precipitation for the next 7 days, and historical meteorological sequences for the same period), user characteristic data (e.g., user industry classification, electricity consumption habit labels such as residential peak-valley electricity consumption type and industrial stable production type), and power grid status data, such as the output curves of distributed photovoltaic / energy storage, charging and discharging records, and periods of reverse power flow.
[0054] In step two, the standardized value is calculated using the following formula: Standardized value = (Original value - Mean) / Standard deviation; where the mean and standard deviation are calculated based on three years of historical data; ensuring that the newly collected data is consistent with the distribution of historical data, laying the foundation for subsequent fusion and repair.
[0055] In step three, the time-series decomposition algorithm is specifically: Load(Xi) = Trend(Xi) + Seasonal(Xi) + Residual(Xi);
[0056] Trend component: reflects the long-term growth / decline trend of load (e.g., 5% annual growth of residential load); Seasonal component: reflects intraday and weekly periodicity (e.g., double peaks of residential load at 8 am and 7 pm); Residual component: the fluctuation part affected by sudden factors (e.g., extreme weather, equipment failure).
[0057] In step five, the user load characteristics include the daily load peak-valley difference, weekly average load fluctuation rate, industry electricity consumption ratio, and holiday load attenuation rate; the power supply impact characteristics include the photovoltaic output forecast (based on meteorological radiation data), energy storage charging and discharging status (charging / discharging / standby), and the probability of reverse power flow (combined with historical time period statistics); the correlation characteristics include using fast Fourier transform to convert the time-domain load into the frequency domain and extracting high-frequency correlation components with temperature and photovoltaic output.
[0058] In step six, the weighted fusion formula is as follows: Fused_Feature = 0.4 × load characteristic + 0.3 × power supply characteristic + 0.2 × meteorological characteristic + 0.1 × user characteristic.
[0059] In step eight, the combined loss function is used: Loss = MSE + 0.1 × MAE; MSE optimizes the overall accuracy, and MAE reduces extreme value errors.
[0060] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or basic characteristics. Therefore, the embodiments should be considered exemplary and non-limiting in all respects. The scope of the invention is defined by the appended claims rather than the foregoing description. Therefore, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention, and no reference numerals in the claims should be construed as limiting the scope of the claims.
[0061] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A short-term load forecasting method for a power distribution network, characterized by, The specific step flow of the short-term load forecasting method for the power distribution network is as follows: Step one, core data classification collection: classify and collect load-related data and influencing factor data; Step two, data standardization processing: use Z-score standardization to unify the scale of different dimension data; Step three, load data component decomposition: decompose the load data into three parts through a time series decomposition algorithm; Step four, accurate repair of each component: judge the type of data problem, and determine which type it is among short-term missing, long-term missing, reverse flow anomaly and random abnormal value; Step five, feature extraction dimension: extract the feature dimension of user load characteristics, power supply influence characteristics and correlation characteristics; Step six, feature screening and fusion: calculate the correlation between the features and the net load through mutual information entropy, retain the features with a correlation greater than or equal to 0.6, integrate multiple source features using a weighted fusion formula, and adjust the weight based on historical prediction error; Step seven, model structure design: design the model structure as input layer, coding layer, decoding layer and output layer; the input layer receives the fused feature sequence; the coding layer models the inter-substation load correlation through GCN to extract spatial correlation features; The decoding layer uses the Transformer multi-head self-attention mechanism to capture time sequence dependence, and the output layer uses a full connection layer to output 24 / 48 / 72-hour 15-minute load prediction values; Step eight, model training and optimization: divide the historical data into training set, validation set and test set according to the ratio of 7:2:1, add artificially constructed missing data samples to enhance robustness, use a combination loss function, and perform iterative training with an initial learning rate of 0.001, a decay of 10% every 100 rounds, and a stop when the validation set error does not decrease for five consecutive rounds; Step nine, multi-period rolling prediction: at 6 am every day, predict the load from 7 am to 7 am the next day based on the latest 24-hour data; update every 4 hours, adjust the subsequent prediction results combined with real-time load and weather correction values, and trigger adaptive correction when the deviation between real-time load and prediction value is greater than or equal to 10%; Step ten, prediction result output: output the basic output and auxiliary output of the prediction result; the basic output is the 15-minute load prediction value and prediction error interval for the next 1-7 days; The auxiliary output is the load peak value occurrence period and distributed power consumption suggestion; During the result output process, when the predicted load exceeds 80% of the line carrying capacity, an overload warning is triggered and a load transfer suggestion is pushed.
2. The method for short term load forecasting for an electric distribution network according to claim 1, characterized in that: In step one, the load-related data includes historical load data and real-time load data of the target substation / region collected through the power distribution terminal, as well as the timestamps and fault types of meter faults and communication interruptions; The influencing factor data includes weather data, user characteristic data and power grid state data.
3. The method for short term load forecasting for an electric distribution network according to claim 1, characterized in that: In step two, the calculation formula of the standardized value is as follows: standardized value = (original value - mean) / standard deviation; wherein the mean and standard deviation are calculated based on historical data for three years.
4. The method for short term load forecasting for an electric distribution network according to claim 1, characterized in that: In step three, the time series decomposition algorithm is as follows: Load(Xi) = Trend(Xi) + Seasonal(Xi) + Residual(Xi); Trend: reflects the long-term growth / downtrend of load; Seasonal: embodies the daily / weekly periodicity; Residual: the fluctuation part affected by sudden factors.
5. The method for short term load forecasting for an electric distribution network according to claim 1, characterized in that: In the step five, the user load characteristics include daily load peak-valley difference, weekly load fluctuation rate, industry electricity consumption proportion, and holiday load attenuation rate; the power supply influence characteristics include photovoltaic output prediction value, energy storage charging and discharging state, and reverse power flow occurrence probability; the correlation characteristics include converting the time-domain load into the frequency domain by using fast Fourier transform, and extracting high-frequency correlation components of temperature and photovoltaic output.
6. The method for short term load forecasting for an electric distribution network according to claim 1, characterized in that: In the step six, the weighted fusion formula is specifically: Fused_Feature = 0.4 x load characteristics + 0.3 x power supply characteristics + 0.2 x meteorological characteristics + 0.1 x user characteristics.
7. The method for short term load forecasting for an electric distribution network according to claim 1, characterized in that: In the step eight, the combined loss function is: Loss = MSE + 0.1 x MAE.