An adaptive hybrid power load probability prediction method and device
By combining discrete wavelet transform, stacked bidirectional LSTM network and extreme gradient boosting algorithm, the problem of multi-scale features and residual heterogeneity in load forecasting is solved, achieving high-precision load forecasting and improving the dispatch accuracy and security of power system.
Patent Information
- Application Number
- CN202610988445.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-03
- Publication Date
- 2026-11-03
AI Technical Summary
Existing technologies struggle to balance the multi-scale time-frequency characteristics and residual structural properties of loads in high-penetration distributed energy environments, resulting in insufficient load forecasting accuracy and impacting the accuracy and security of power system dispatch.
Discrete wavelet transform is used to extract multi-scale time-frequency features. Combined with stacked bidirectional long short-term memory network and time-series pattern attention mechanism, the difference between the preliminary predicted value and the true value is calculated as the target residual sequence to be fitted. The extreme gradient boosting algorithm is used for fitting to generate residual compensation value. Finally, the final prediction result is output by superimposing the values.
It significantly improves the accuracy and robustness of day-ahead load forecasting, enhances adaptability under strong fluctuation conditions, and provides reliable dispatch support for the power system.
Smart Images
Figure CN123030858A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of power load forecasting, and in particular to an adaptive hybrid power load probability forecasting method, apparatus, equipment, and storage medium. Background Technology
[0002] With the increasing penetration of distributed energy resources, such as photovoltaic and wind power, in power distribution networks, the power load curve exhibits significant uncertainty and drastic fluctuations. In traditional load structures, user electricity consumption behavior is relatively regular, and load timing exhibits strong periodicity and stability. However, distributed power generation output is affected by random factors such as weather conditions and seasonal variations, resulting in intermittent power output and anti-peak-shaving characteristics. This leads to a significant increase in the amplitude and rate of change of net load fluctuations, and a significant enhancement in the non-stationarity and nonlinearity of load timing. These complex characteristics pose a severe challenge to short-term load forecasting.
[0003] Accurate short-term load forecasting, especially day-ahead load forecasting, is a crucial foundation for power system operation and dispatch. In the day-ahead market, forecast results are key to formulating unit start-up and shutdown plans, determining generation output curves, allocating reserve capacity, and assessing grid safety margins. Insufficient forecast accuracy will lead to suboptimal unit combination decisions, resulting in frequent adjustments or start-ups and shutdowns of conventional units, increasing coal consumption and mechanical wear. Simultaneously, deviations in dispatch plans will force the system to utilize more spinning reserves, reducing economic operating efficiency and potentially causing power flow exceeding limits or voltage stability issues, threatening the safe and reliable operation of the power grid.
[0004] Currently, commonly used forecasting methods include traditional statistical learning models such as time series analysis and support vector regression, as well as deep learning models based on recurrent neural networks and long short-term memory networks. The former has limited ability to fit strongly nonlinear relationships and is difficult to capture abrupt changes under highly volatile loads; the latter, although it can extract time-series dependencies, often ignores the decomposition characteristics of the load series at different time scales and lacks an effective compensation mechanism for irregular components in the forecast residuals, resulting in large forecast biases during periods of extreme volatility.
[0005] Therefore, there is an urgent need for a forecasting method that can fully exploit the multi-scale time-frequency characteristics of loads and effectively correct residual biases, so as to improve the accuracy and robustness of day-ahead load forecasting in a high-penetration distributed energy environment and provide reliable support for the economic dispatch and safe operation of the power system. Summary of the Invention
[0006] The present invention aims to at least partially solve one of the technical problems in the related art.
[0007] To address this issue, this invention discloses an adaptive hybrid power load probabilistic prediction method. This method uses discrete wavelet transform on the original load time-series data to decompose low-frequency trends and high-frequency details, generating a multi-scale time-frequency feature set. The feature set, concatenated with external influencing factors, is input into a stacked bidirectional long short-term memory network to extract bidirectional time-series dependencies. These dependencies are then processed through convolutional filtering and weighted fusion using a time-series pattern attention mechanism to generate preliminary predicted values. The difference between the preliminary predicted values and the actual values is calculated as a residual sequence. Using the feature set and the preliminary predicted values as input, an extreme gradient boosting algorithm is used to fit the residuals to generate compensation values. Finally, the preliminary predicted values and compensation values are superimposed to output the final result. This invention solves the problem that conventional prediction methods struggle to simultaneously consider multi-scale features and residual structural variations, achieving high-precision adaptive load prediction.
[0008] Another objective of this invention is to provide an adaptive hybrid power load probability prediction device.
[0009] The third objective of this invention is to provide a computer device.
[0010] A fourth objective of this invention is to provide a non-transitory computer-readable storage medium.
[0011] To achieve the above objectives, this invention proposes an adaptive hybrid power load probabilistic prediction method, comprising:
[0012] Discrete wavelet transform is performed on the original power load time series data to decompose it into a low-frequency trend component and multiple high-frequency detail components, which are then combined to generate a multi-scale time-frequency feature set. The multi-scale time-frequency feature set is concatenated with external influencing factors and then input into a multi-layer stacked bidirectional long short-term memory network to extract bidirectional temporal dependencies and generate hidden layer states at each time step. The hidden layer states are then convolved and filtered using a temporal pattern attention mechanism and attention weights are assigned. After weighted fusion, the hidden layer states are integrated with the current hidden layer states to generate preliminary load prediction values. The difference between the preliminary load forecast and the actual value is calculated as the target residual sequence to be fitted. The multi-scale time-frequency feature set and the preliminary load forecast are used as input features. The extreme gradient boosting algorithm is used to fit the target residual sequence to generate residual compensation values. The initial load forecast value is superimposed with the residual compensation value to output the final forecast result.
[0013] An adaptive hybrid power load probability prediction method according to an embodiment of the present invention may also have the following additional technical features: In one embodiment of the present invention, performing discrete wavelet transform on the original power load time-series data includes: Obtain raw power load time-series data as input signal; Using the Daubechies wavelet family as the basis functions and setting the decomposition level to 5, multi-resolution discrete wavelet decomposition is performed on the input signal. The discrete wavelet transform decomposes the input signal into multiple frequency bands, and the decomposition expression is as follows:
[0014] in, For wavelet basis functions, These are the approximation coefficients, i.e., the low-frequency components. and These represent the scale parameter and the translation parameter, respectively. After five levels of decomposition, the fifth level wavelet decomposition expression is obtained:
[0015] The output consists of 6 components, including 1 low-frequency approximation component. and 5 high-frequency detail components to Among them, the low-frequency approximation component Reflecting the long-term macro trend of power load changes, the five high-frequency detail components, in descending order of frequency, reflect instantaneous noise disturbances, minute-level rapid fluctuations, hourly load impacts, half-day pattern changes, and diurnal periodic fluctuations.
[0016] In one embodiment of the present invention, the multi-scale time-frequency feature set is concatenated with external influencing factors and then input into a multi-layer stacked bidirectional long short-term memory network, including: A multi-scale time-frequency feature set and collected external meteorological factor data are acquired, and the two are horizontally concatenated according to the same timestamp to form a multi-dimensional input feature matrix containing time-series load features and external environmental features. ; The multidimensional input feature matrix As input, the sequence is sequentially fed into each layer of a multi-layered stacked bidirectional long short-term memory network. Within each layer, the sequence is propagated forward along the forward temporal direction to extract historical information, and simultaneously propagated backward along the backward temporal direction to extract future information. The mathematical model of the bidirectional long short-term memory network is as follows:
[0017] in, Indicates time step Input, This indicates that the hidden layer of the feedforward LSTM is at time step The output, This indicates that the backward LSTM hidden layer is at time step The output, This represents the bidirectional concatenation result of the hidden layer output; The information is passed layer by layer, with the top layer outputting the final hidden layer state information at each time step, which is then combined to form a hidden layer information matrix. ;in Indicates time step Hidden layer information, This indicates the preset time sliding window length.
[0018] In one embodiment of the present invention, a temporal pattern attention mechanism is used to perform convolutional filtering on the hidden layer state and assign attention weights, including: Obtain the final hidden layer state information at each time step, and slide the window according to the preset time length. Extract the hidden layer states from multiple consecutive time steps prior to the current time step to form a hidden layer state matrix. ; Employing multiple preset one-dimensional convolutional filters Along the time dimension, the hidden layer state matrix Each row of variables in the matrix undergoes a sliding convolution operation to extract the local temporal variation pattern of each variable within different time windows, generating a temporal pattern feature matrix. The convolution operation expression is:
[0019] in, Indicates a filter. Indicates the first Row vectors are filtered Extraction length is The result after the features, This represents the maximum size of the weights obtained from the filter, and its value is... ; Calculate the hidden layer function at the current time step With time-series pattern feature matrix The expression for calculating the association strength and attention weights of each row of feature vectors is as follows:
[0020]
[0021] in, express The Middle The vector of rows, for The weight coefficient matrix, To be allocated to the first Attention weights for row vectors; The attention weight With time-series pattern feature matrix The corresponding row vectors are multiplied element-wise with weights, and the weighted results of each row are summed and aggregated. The weighted summation expression is:
[0022] in, Represents the timing effect vector. Indicates the dimension of the input variable.
[0023] In one embodiment of the present invention, integrating the weighted fusion result with the current hidden layer state to generate a preliminary load prediction value includes: Obtaining the temporal influence vector Hidden layer function at the current time step ; Translate the temporal effect vector Hidden layer function at the current time step Perform linear fusion; the fusion expression is:
[0024] in, and For matrix parameter weights, This is the final hidden layer output obtained after fusion; Output the final hidden layer The input is fed into a fully connected mapping layer, which performs a linear transformation and outputs a single numerical value as the preliminary load forecast for the current time step. .
[0025] In one embodiment of the present invention, using a multi-scale time-frequency feature set and preliminary load forecast values as input features, the target residual sequence is fitted using an extreme gradient boosting algorithm, including: Obtain multi-scale time-frequency feature sets and preliminary load forecasts. The two are horizontally concatenated according to the same timestamp to construct the residual correction input feature matrix; Obtain from preliminary load forecast values Compared with the true value The target residual sequence obtained by subtracting time points one by one The residual calculation expression is:
[0026] The target residual sequence As the fitting objective of the extreme gradient boosting algorithm, the residual correction input feature matrix is input into the extreme gradient boosting algorithm. The objective function of the extreme gradient boosting algorithm is to minimize the residual loss, expressed as:
[0027] in, The squared error loss function is for the residuals. This is a regularization term used to penalize model complexity. Indicates the first in the integration A decision tree; The extreme gradient boosting algorithm follows a sequential ensemble strategy, adding weak learners to decision trees one by one. Each new decision tree is trained using the prediction residuals of all previous trees as the fitting object. The model parameters are iteratively updated by minimizing the residual loss function, gradually approximating the nonlinear fluctuation pattern in the residual sequence. After all decision trees are integrated, the residual correction input feature matrix is input into the trained extreme gradient boosting model. The model is then weighted and aggregated by voting from each decision tree to output the residual compensation prediction value. .
[0028] In one embodiment of the present invention, the final prediction result is output by superimposing the preliminary load forecast value and the residual compensation value, including: Obtain preliminary load forecasts at the same forecast time. With residual compensation prediction value ; Preliminary load forecast With residual compensation prediction value The expression for numerical addition is:
[0029] in, To achieve the final prediction result, the residual compensation value is used to correct the residual portion of the initial load forecast that was not captured by the long short-term memory network point by point. Output the numerical value after residual correction. This serves as the final adaptive hybrid power load probability prediction result.
[0030] To achieve the above objectives, another aspect of the present invention proposes an adaptive hybrid power load probability prediction device, comprising: The time-frequency decomposition module is used to perform discrete wavelet transform on the original power load time-series data, decomposing it into a low-frequency trend component and multiple high-frequency detail components, which are then combined to generate a multi-scale time-frequency feature set. The time-series prediction module is used to concatenate multi-scale time-frequency feature sets with external influencing factors and input them into a multi-layer stacked bidirectional long short-term memory network to extract bidirectional temporal dependencies and generate hidden layer states at each time step. The hidden layer states are convolutionally filtered and weighted using a time-series pattern attention mechanism. After weighted fusion, the hidden layer states are integrated with the current hidden layer states to generate preliminary load prediction values. The residual compensation module is used to calculate the difference between the preliminary load forecast value and the actual value as the target residual sequence to be fitted. It uses a multi-scale time-frequency feature set and the preliminary load forecast value as input features, and uses the extreme gradient boosting algorithm to fit the target residual sequence to generate residual compensation values. The fusion output module is used to superimpose the preliminary load forecast value and the residual compensation value to output the final forecast result.
[0031] This invention discloses an adaptive hybrid power load probabilistic prediction method and apparatus. It extracts multi-scale time-frequency features of the load through discrete wavelet transform, extracts bidirectional time-series dependencies using a stacked bidirectional long short-term memory network and a time-series pattern attention mechanism, and generates preliminary prediction values. Then, an extreme gradient boosting algorithm is used to nonlinearly fit and compensate for the prediction residuals, achieving organic synergy between time-series decomposition, deep modeling, and residual correction. This method effectively solves the prediction bias problem caused by the difficulty of existing technologies in simultaneously considering the multi-scale features of the load sequence and the structural differences in residuals. It significantly improves the accuracy and robustness of day-ahead load prediction in high-penetration distributed energy access scenarios, enhances the adaptability and engineering practical value of the prediction method under strong fluctuation conditions, and provides reliable data support for day-ahead dispatch decisions in power systems.
[0032] To achieve the above objectives, a third aspect of this application provides a computer device comprising a processor and a memory; wherein the processor runs a program corresponding to the executable program code by reading executable program code stored in the memory, for implementing an adaptive hybrid power load probability prediction method as described in the first aspect embodiment.
[0033] To achieve the above objectives, a fourth aspect of this application provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements an adaptive hybrid power load probability prediction method as described in the first aspect embodiment.
[0034] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description
[0035] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 This is a flowchart of an adaptive hybrid power load probability prediction method according to an embodiment of the present invention; Figure 2 This is a structural diagram of a multilayer stacked bidirectional LSTM according to an embodiment of the present invention; Figure 3This is a structural diagram of the TPA according to an embodiment of the present invention; Figure 4 This is a flowchart of TPA-multilayer stacked bidirectional LSTM according to an embodiment of the present invention; Figure 5 This is an overall flowchart according to an embodiment of the present invention; Figure 6 This is a schematic diagram of a device for predicting the probability of adaptive hybrid power load according to an embodiment of the present invention; Figure 7 It is a computer device according to an embodiment of the present invention. Detailed Implementation
[0036] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.
[0037] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. 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 should fall within the scope of protection of the present invention.
[0038] The following description, with reference to the accompanying drawings, describes an adaptive hybrid power load probability prediction method, apparatus, device, and storage medium based on embodiments of the present invention.
[0039] The core idea of this invention is to address the problem that existing prediction methods, under the high penetration of distributed energy access, exhibit strong uncertainty and non-stationarity in power load curves, and are unable to simultaneously consider multi-scale time-frequency characteristics and residual anomalies, leading to insufficient accuracy. This is achieved by constructing an adaptive hybrid load prediction method covering the entire process of "time-frequency decomposition—time-series modeling—residual compensation." First, the original load time-series data is subjected to discrete wavelet transform, decomposing it into low-frequency trends and high-frequency details to generate a multi-scale feature set. Based on this, the feature set is concatenated with external influencing factors and input into a multi-layer stacked bidirectional long short-term memory network. A time-series pattern attention mechanism is used to extract bidirectional time-series dependencies and perform weighted fusion to generate preliminary prediction values. The difference between the preliminary prediction value and the actual value is calculated as a residual sequence. Using the feature set and the preliminary prediction value as input, an extreme gradient boosting algorithm is used to perform nonlinear fitting on the residuals to generate residual compensation values. Finally, the preliminary prediction value and the compensation value are superimposed to output the final prediction result. By integrating multi-scale decomposition, bidirectional time series extraction, and adaptive residual correction, this invention effectively solves the problem of large prediction deviations in traditional methods under strong fluctuation conditions, significantly improves the accuracy and robustness of day-ahead load forecasting, and enhances its engineering practical value in high-penetration distributed energy scenarios.
[0040] The following description, with reference to the accompanying drawings, describes an adaptive hybrid power load probability prediction method and apparatus according to an embodiment of the present invention.
[0041] like Figure 1 As shown, the adaptive hybrid power load probability prediction method of the present invention includes the following steps: S1 performs discrete wavelet transform on the original power load time series data, decomposing it into a low-frequency trend component and multiple high-frequency detail components, which are then combined to generate a multi-scale time-frequency feature set.
[0042] In one embodiment of the present invention, it is set that This represents the original time series. Discrete Wavelet Transform (DWT) will... It is decomposed into multiple frequency bands, thus providing its representation in both the time and frequency domains simultaneously. The discrete wavelet transform can be expressed as: (1) in, For wavelet basis functions, The approximation coefficients (low-frequency components), and and Let represent the scale parameter and translation parameter in formula (1), respectively.
[0043] In this invention, wavelet transform is used to decompose the load data into components of different frequencies. Specifically, the Daubechies wavelet (dbN) is used, with a decomposition level of 5, to generate approximation coefficients. With detail coefficient This decomposition yields an approximate component reflecting the long-term trend. And five detailed components for capturing short-term changes. to These detailed components sequentially represent changes in frequency that gradually decrease: among which The highest frequency change corresponds to (e.g., noise or rapid fluctuations), while This represents localized changes at lower frequencies. These components help the model identify and learn patterns associated with mutations, potential anomalies (such as "unpredictable events"), and seasonal effects.
[0044] The fifth-level wavelet decomposition is shown in formula (2): (2) These coefficients are then used as input features in the LSTM-XGBoost hybrid model. The physical meaning of each coefficient is shown in Table 1.
[0045] Table 1
[0046] S2, after concatenating the multi-scale time-frequency feature set with external influencing factors, input it into a multi-layer stacked bidirectional long short-term memory network to extract bidirectional temporal dependencies and generate hidden layer states at each time step; use a temporal pattern attention mechanism to perform convolutional filtering on the hidden layer states and assign attention weights, then integrate the weighted fusion with the current hidden layer state to map and generate preliminary load prediction values.
[0047] In one embodiment of the present invention, the present invention utilizes a bidirectional LSTM network to capture temporal information of the input sequence from past and future time steps for forward and backward training. The mathematical model of the bidirectional LSTM neural network is shown in Equation (3).
[0048] (3) in, Indicates time step Input; This indicates that the hidden layer of the feedforward LSTM is at time step The output; This indicates that the backward LSTM hidden layer is at time step The output; This represents the output of the hidden layer.
[0049] Multi-layer stacked bidirectional LSTM is a process of continuously expanding the depth of a bidirectional LSTM network, in which the neural network undergoes multiple training iterations to improve prediction accuracy. The structure of a multi-layer stacked bidirectional LSTM is as follows: Figure 2 As shown.
[0050] Traditional attention mechanisms primarily focus on extracting information relevant to the current time step, which works well for single time series data. However, in power load forecasting, each time step contains multiple variables, including power time series and meteorological factors. Due to the diverse and complex nonlinear relationships among these variables, traditional attention mechanisms, which average information from multiple time steps, struggle to effectively capture the importance of specific time steps.
[0051] The Time-Based Approach (TPA) mechanism utilizes the frequency domain information of time series data to extract inherent time characteristics. It effectively reveals the intrinsic relationship between various variables and the electricity load time series. The structure of TPA is as follows: Figure 3 As shown. Among them, h This represents the output information of the hidden layer of the previous level neural network, with each output representing the temporal state of the previous level network. Subsequently, multiple convolutional layer filters are used to extract features.
[0052] Assuming the hidden layer information generated by a multi-layer stacked bidirectional LSTM network is Among them, Indicates time step Hidden layer information, This indicates the length of the time sliding window used to select a certain length of time data. This represents the maximum length of the weights extracted by the filter, and is usually equal to the sliding window length. For the original hidden layer information Perform convolution to generate an n k-dimensional matrix .
[0053] (4) in, Indicates a filter. Indicates the first Row vectors are filtered Extraction length is The result after considering the features. middle, Represents a row vector. Indicates time, This indicates the length of the filter, while T This represents the maximum size of the weights obtained from the filter, typically taking the value [value to be filled in]. ω .
[0054] Weight By calculating the hidden layer function and "to obtain" means Each row in the table corresponds to the hidden layer function. The intensity of the impact. Weighting coefficients The expressions are shown in formulas (5) and (6) below.
[0055] (5) (6) in, express The Middle A vector of rows.
[0056] right Perform a weighted summation, where express All rows in The cumulative effect, i.e., the sequential effect.
[0057] (7) in, Indicates the dimension of the input variable.
[0058] Finally, considering the effects of time series The overall effect yields the final predicted value. As shown below: (8) in, and These are the matrix parameter weights.
[0059] The flowchart of TPA-multilayer stacked bidirectional LSTM is as follows: Figure 4 As shown.
[0060] S3, calculate the difference between the preliminary load forecast and the actual value as the target residual sequence to be fitted, use the multi-scale time-frequency feature set and the preliminary load forecast as input features, and use the extreme gradient boosting algorithm to fit the target residual sequence to generate residual compensation values.
[0061] In one embodiment of the present invention, after obtaining the prediction values of a multi-layer stacked bidirectional LSTM, the residual error... rt Calculated using the following formula: (9) in, for The actual load value at any given time. For LSTM in tThe predicted values at time points are shown in Equation (9). These residuals... This represents the portion of data that LSTM failed to capture, typically corresponding to unpredictable events and short-term fluctuations.
[0062] XGBoost is used to model and correct these residuals. XGBoost fits the residual error by sequentially adding weak learners (decision trees). Its objective function is to minimize the residual loss in formula (10), as shown below: (10) in, This is the loss function (squared error) for the residuals. This is a regularization term used to penalize model complexity. Indicates the first in the integration A decision tree.
[0063] XGBoost parameters (such as max_depth, learning_rate, n_estimators) are optimized using RandomizedSearchCV to further improve model performance.
[0064] S4 superimposes the preliminary load forecast value with the residual compensation value and outputs the final forecast result.
[0065] In one embodiment of the present invention, the overall process is as follows: Figure 5 As shown. The final predicted value of the mixture model. By using the prediction values of LSTM Correction value with XGBoost The result obtained by combining these formulas is shown in formula (11): (11).
[0066] The method of this invention effectively solves the problem of prediction bias caused by the difficulty of taking into account the multi-scale fluctuation characteristics of load and residual structural differences in the prior art. It significantly improves the accuracy and robustness of day-ahead load prediction in high-penetration distributed energy access scenarios and enhances the adaptability and engineering practical value of the prediction method under highly uncertain operating conditions.
[0067] To achieve the above invention, such as Figure 6 As shown, this embodiment also provides an adaptive hybrid power load probability prediction device 10, which includes: The time-frequency decomposition module 100 is used to perform discrete wavelet transform on the original power load time series data, decompose it into a low-frequency trend component and multiple high-frequency detail components, and combine them to generate a multi-scale time-frequency feature set.
[0068] The time-series prediction module 200 is used to concatenate multi-scale time-frequency feature sets with external influencing factors and input them into a multi-layer stacked bidirectional long short-term memory network to extract bidirectional temporal dependencies and generate hidden layer states at each time step. The hidden layer states are convolutionally filtered and attention weights are assigned using a time-series pattern attention mechanism. After weighted fusion, the hidden layer states are integrated with the current hidden layer states to map and generate preliminary load prediction values.
[0069] The residual compensation module 300 is used to calculate the difference between the preliminary load forecast value and the actual value as the target residual sequence to be fitted. It uses a multi-scale time-frequency feature set and the preliminary load forecast value as input features, and uses an extreme gradient boosting algorithm to fit the target residual sequence to generate residual compensation values.
[0070] The fusion output module 400 is used to superimpose the preliminary load forecast value and the residual compensation value to output the final forecast result.
[0071] The apparatus of this invention achieves integrated collaboration across the entire process of time-frequency decomposition, bidirectional time-series modeling, and adaptive residual correction. This apparatus effectively solves the problem of prediction bias caused by the difficulty of existing technologies in simultaneously addressing the multi-scale fluctuation characteristics of load and the structural nature of residuals. It significantly improves the accuracy and robustness of day-ahead load forecasting in high-penetration distributed energy scenarios, and enhances the adaptability and engineering practical value of the forecasting method under highly uncertain operating conditions.
[0072] To implement the methods of the above embodiments, the present invention also provides a computer device, such as... Figure 7 As shown, the computer device 600 includes a memory 601 and a processor 602; wherein, the processor 602 reads the executable program code stored in the memory 601 to run a program corresponding to the executable program code, so as to implement the various steps of the adaptive hybrid power load probability prediction method described above.
[0073] To implement the above embodiments, this application also proposes a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements an adaptive hybrid power load probability prediction method as described in the foregoing embodiments.
[0074] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.
[0075] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.
Claims
1. A probabilistic prediction method for adaptive hybrid power load, characterized in that, include: Discrete wavelet transform is performed on the original power load time series data to decompose it into a low-frequency trend component and multiple high-frequency detail components, which are then combined to generate a multi-scale time-frequency feature set. The multi-scale time-frequency feature set is concatenated with external influencing factors and then input into a multi-layer stacked bidirectional long short-term memory network to extract bidirectional temporal dependencies and generate hidden layer states at each time step. The hidden layer states are then convolved and filtered using a temporal pattern attention mechanism and attention weights are assigned. After weighted fusion, the hidden layer states are integrated with the current hidden layer states to generate preliminary load prediction values. The difference between the preliminary load forecast and the actual value is calculated as the target residual sequence to be fitted. The multi-scale time-frequency feature set and the preliminary load forecast are used as input features. The extreme gradient boosting algorithm is used to fit the target residual sequence to generate residual compensation values. The initial load forecast value is superimposed with the residual compensation value to output the final forecast result.
2. The method as described in claim 1, characterized in that, Discrete wavelet transform is performed on the original power load time series data, including: Obtain raw power load time-series data as input signal; Using the Daubechies wavelet family as the basis functions and setting the decomposition level to 5, multi-resolution discrete wavelet decomposition is performed on the input signal. The discrete wavelet transform decomposes the input signal into multiple frequency bands, and the decomposition expression is as follows: in, For wavelet basis functions, These are the approximation coefficients, i.e., the low-frequency components. and These represent the scale parameter and the translation parameter, respectively. After five levels of decomposition, the fifth level wavelet decomposition expression is obtained: The output consists of 6 components, including 1 low-frequency approximation component. and 5 high-frequency detail components to Among them, the low-frequency approximation component Reflecting the long-term macro trend of power load changes, the five high-frequency detail components, in descending order of frequency, reflect instantaneous noise disturbances, minute-level rapid fluctuations, hourly load impacts, half-day pattern changes, and diurnal periodic fluctuations.
3. The method as described in claim 1, characterized in that, The multi-scale time-frequency feature set is concatenated with external influencing factors and then input into a multi-layer stacked bidirectional long short-term memory network, including: A multi-scale time-frequency feature set and collected external meteorological factor data are acquired, and the two are horizontally concatenated according to the same timestamp to form a multi-dimensional input feature matrix containing time-series load features and external environmental features. ; The multidimensional input feature matrix As input, the sequence is sequentially fed into each layer of a multi-layered stacked bidirectional long short-term memory network. Within each layer, the sequence is propagated forward along the forward temporal direction to extract historical information, and simultaneously propagated backward along the backward temporal direction to extract future information. The mathematical model of the bidirectional long short-term memory network is as follows: in, Indicates time step Input, This indicates that the hidden layer of the feedforward LSTM is at time step The output, This indicates that the backward LSTM hidden layer is at time step The output, This represents the bidirectional concatenation result of the hidden layer output; The information is passed layer by layer, with the top layer outputting the final hidden layer state information at each time step, which is then combined to form a hidden layer information matrix. ;in Indicates time step Hidden layer information, This indicates the preset time sliding window length.
4. The method as described in claim 1, characterized in that, A temporal attention mechanism is used to perform convolutional filtering on the hidden layer states and assign attention weights, including: Obtain the final hidden layer state information at each time step, and slide the window according to the preset time length. Extract the hidden layer states from multiple consecutive time steps prior to the current time step to form a hidden layer state matrix. ; Employing multiple preset one-dimensional convolutional filters Along the time dimension, the hidden layer state matrix Each row of variables in the matrix undergoes a sliding convolution operation to extract the local temporal variation pattern of each variable within different time windows, generating a temporal pattern feature matrix. The convolution operation expression is: in, Indicates a filter. Indicates the first Row vectors are filtered Extraction length is The result after the features, This represents the maximum size of the weights obtained from the filter, and its value is... ; Calculate the hidden layer function at the current time step With time-series pattern feature matrix The expression for calculating the association strength and attention weights of each row of feature vectors is as follows: in, express The Middle The vector of rows, for The weight coefficient matrix, To be allocated to the first Attention weights for row vectors; The attention weight With time-series pattern feature matrix The corresponding row vectors are multiplied element-wise with weights, and the weighted results of each row are summed and aggregated. The weighted summation expression is: in, Represents the timing effect vector. Indicates the dimension of the input variable.
5. The method as described in claim 1, characterized in that, The weighted fusion result is integrated and mapped with the current hidden layer state to generate preliminary load forecast values, including: Obtaining the temporal influence vector Hidden layer function at the current time step ; Translate the temporal effect vector Hidden layer function at the current time step Perform linear fusion; the fusion expression is: in, and For matrix parameter weights, This is the final hidden layer output obtained after fusion; Output the final hidden layer The input is fed into a fully connected mapping layer, which performs a linear transformation and outputs a single numerical value as the preliminary load forecast for the current time step. .
6. The method as described in claim 1, characterized in that, Using a multi-scale time-frequency feature set and preliminary load forecasts as input features, the target residual sequence is fitted using an extreme gradient boosting algorithm, including: Obtain multi-scale time-frequency feature sets and preliminary load forecasts. The two are horizontally concatenated according to the same timestamp to construct the residual correction input feature matrix; Obtain from preliminary load forecast values Compared with the true value The target residual sequence obtained by subtracting time points one by one The residual calculation expression is: The target residual sequence As the fitting objective of the extreme gradient boosting algorithm, the residual correction input feature matrix is input into the extreme gradient boosting algorithm. The objective function of the extreme gradient boosting algorithm is to minimize the residual loss, expressed as: in, The squared error loss function is for the residuals. This is a regularization term used to penalize model complexity. Indicates the first in the integration A decision tree; The extreme gradient boosting algorithm follows a sequential ensemble strategy, adding weak learners to decision trees one by one. Each new decision tree is trained using the prediction residuals of all previous trees as the fitting object. The model parameters are iteratively updated by minimizing the residual loss function, gradually approximating the nonlinear fluctuation pattern in the residual sequence. After all decision trees are integrated, the residual correction input feature matrix is input into the trained extreme gradient boosting model. The model is then weighted and aggregated by voting from each decision tree to output the residual compensation prediction value. .
7. The method according to claim 1, characterized in that, The final forecast result is output by superimposing the preliminary load forecast value and the residual compensation value, including: Obtain preliminary load forecasts at the same forecast time. With residual compensation prediction value ; Preliminary load forecast With residual compensation prediction value The expression for numerical addition is: in, To achieve the final prediction result, the residual compensation value is used to correct the residual portion of the initial load forecast that was not captured by the long short-term memory network point by point. Output the numerical value after residual correction. This serves as the final adaptive hybrid power load probability prediction result.
8. A device for predicting the probability of adaptive hybrid power load, characterized in that, include: The time-frequency decomposition module is used to perform discrete wavelet transform on the original power load time-series data, decomposing it into a low-frequency trend component and multiple high-frequency detail components, which are then combined to generate a multi-scale time-frequency feature set. The time-series prediction module is used to concatenate multi-scale time-frequency feature sets with external influencing factors and input them into a multi-layer stacked bidirectional long short-term memory network to extract bidirectional temporal dependencies and generate hidden layer states at each time step. The hidden layer states are convolutionally filtered and weighted using a time-series pattern attention mechanism. After weighted fusion, the hidden layer states are integrated with the current hidden layer states to generate preliminary load prediction values. The residual compensation module is used to calculate the difference between the preliminary load forecast value and the actual value as the target residual sequence to be fitted. It uses a multi-scale time-frequency feature set and the preliminary load forecast value as input features, and uses the extreme gradient boosting algorithm to fit the target residual sequence to generate residual compensation values. The fusion output module is used to superimpose the preliminary load forecast value and the residual compensation value to output the final forecast result.
9. An electronic device, characterized in that, include: processor; The memory stores executable instructions; when the processor executes the instructions, it implements the adaptive hybrid power load probability prediction method as described in any one of claims 1-7.
10. A computer-readable storage medium, characterized in that, The device contains a computer program that, when executed by a processor, implements an adaptive hybrid power load probability prediction method as described in any one of claims 1-7.