Power system net load prediction method based on regular decomposition and double-branch prediction

Abstract

The invention discloses a power system net load prediction method based on regular decomposition and double-branch prediction, and the method comprises the steps: obtaining a historical net load sequence of a target power system and corresponding environment parameters, constructing a target function fusing fitting precision and trend smoothness through employing a regularization optimization method, extracting a long-term trend sequence of a net load, and carrying out the calculation of the long-term trend sequence. And a short-term disturbance sequence is separated. Constructing a trend prediction sub-network based on series connection of a Transform encoder and a long-short-term memory network, and learning a trend evolution rule; meanwhile, a regression prediction sub-network based on environmental parameters is constructed, and a nonlinear mapping relation between disturbance and environmental factors is modeled. And utilizing the two types of sub-networks to respectively predict future trend and disturbance components and superpose the future trend and disturbance components to obtain a multi-time-step net load prediction result. According to the method, the problem that a traditional model is insufficient in trend and disturbance modeling capacity is effectively solved, and the accuracy and stability of load prediction are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of net load prediction, and in particular relates to a method for predicting net load of an electric power system based on regularized decomposition and dual-branch prediction. Background Art

[0002] As the scale of the power system continues to expand and the proportion of renewable energy generation gradually increases, accurately predicting net load has become a critical component in ensuring the safe and stable operation of the power grid. Traditional power load forecasting methods often use time series models or deep learning models to directly predict the overall historical net load series. These methods can capture the time series characteristics and some nonlinear relationships of load data to a certain extent. However, with the increasing proportion of renewable energy access and the increasingly complex load-side response behavior, power load forecasting is facing significant challenges. Net load, as an important variable reflecting the difference between traditional load and renewable energy output, not only exhibits obvious seasonality, daily periodicity, and trend characteristics, but also has a large number of non-stationary fluctuations caused by external disturbances such as weather and holidays. Direct predictions using a single model often fail to account for both long-term trends and short-term fluctuations, resulting in limited prediction accuracy.

[0003] In the prior art, Chinese patent CN115423140A discloses a method for short-term load forecasting of a power system, comprising: obtaining historical load data of the power system and preprocessing the data; optimizing the decomposition mode number and penalty factor of the variational mode decomposition algorithm by the Sky Eagle optimizer algorithm, and decomposing the load data into intrinsic mode function components of different center frequencies, including low-frequency intrinsic mode function components and high-frequency intrinsic mode function components; optimizing the LSSVM prediction models based on the linear kernel function and the radial basis kernel function by the Fireworks algorithm, respectively, predicting the low-frequency intrinsic mode function components and the high-frequency intrinsic mode function components, respectively, obtaining the predicted values ​​of the low-frequency intrinsic mode function components and the high-frequency intrinsic mode function components, and obtaining the final load forecast result.

[0004] However, this method has the following limitations: the decomposition method is highly dependent on preset decomposition parameters, such as the number of modes and penalty factor. Parameter adjustment is complex and lacks a unified standard, resulting in insufficient stability in the decomposition results and difficulty adapting to the diversity of load sequences in different scenarios. Secondly, while variational modal decomposition can decompose the signal into several modal components, it suffers from modal aliasing, which leads to unclear boundaries between trend and disturbance components. The trend component may contain short-term fluctuation noise, affecting the accuracy of subsequent prediction models. Furthermore, this method does not fully consider the impact of environmental factors on the disturbance sequence. The disturbance component is treated as pure noise, ignoring its complex relationship with external parameters such as meteorology, which limits its effective prediction ability for disturbances. Finally, the overall use of a unified model to predict each modal component separately lacks differentiated modeling strategies for trend and disturbance sequences. This results in insufficient representation of the different timescale characteristics of the net load sequence, making it difficult to accurately predict long-term trends and short-term fluctuations.

[0005] Therefore, there is an urgent need for an innovative forecasting method that can accurately extract trend sequences and ensure their smoothness, while combining environmental parameters to perform targeted modeling of disturbance sequences, in order to improve the overall performance of net load forecasting. Summary of the Invention

[0006] The purpose of the present invention is to provide a method for predicting the net load of a power system based on regular decomposition and two-branch prediction in order to overcome the defects of the above-mentioned prior art.

[0007] The purpose of the present invention can be achieved by the following technical solutions:

[0008] The present invention provides a method for predicting net load of a power system based on regularized decomposition and dual-branch prediction, comprising the following steps:

[0009] Obtain the historical net load sequence and corresponding historical environmental parameters of the target power system;

[0010] Based on the regularized optimization method, an objective function is constructed that takes into account the smoothness of the fitting of the historical net load series and the trend series. The trend series of the historical net load series is obtained by solving the objective function.

[0011] Obtaining a disturbance sequence representing short-term rapid fluctuations of the net load according to the trend sequence and the original net load sequence;

[0012] A trend prediction subnetwork based on a series connection between a Transformer encoder and a long short-term memory network is constructed. The trend sequence obtained from the historical net load sequence is used as a sample to train the trend prediction subnetwork.

[0013] For the disturbance sequence, a regression prediction subnetwork based on historical environmental parameters is constructed, and the environmental parameters within the same historical window are used as input to train a disturbance prediction model;

[0014] The trained trend prediction subnetwork and disturbance prediction model are used to predict the trend component and disturbance component of each future time step respectively and superimpose them to obtain the net load prediction value of multiple future time steps.

[0015] Furthermore, the historical environmental parameters include meteorological and renewable energy power generation characteristic parameters related to the net load change, specifically including: temperature, humidity, wind speed, wind direction, and solar irradiance at historical moments.

[0016] Furthermore, the objective function is constructed to take into account the smoothness of the fitting historical net load series and trend series. The objective function is:

[0017]

[0018] Where N is the total number of time points in the historical net load series; x(t) represents the observed value of the historical net load series at the tth time point; y(t) represents the fitted value of the trend series at the tth time point; λ>0 is the trend smoothing regularization coefficient; φ δ (r) is the Huber loss function, which is used to balance the sensitivity of the squared error to normal deviations and the robustness of the linear penalty to mutation points. It is defined as:

[0019]

[0020] Among them, r = x(t) - y(t) is the residual, and δ>0 is the residual threshold, which is used to distinguish normal deviations from abnormal disturbance mutations in the net load sequence.

[0021] Furthermore, obtaining the trend sequence of the historical net load sequence by solving the objective function specifically includes: solving the objective function using a constrained numerical optimization algorithm, wherein the numerical optimization algorithm is a gradient descent method, to obtain the trend sequence that minimizes the objective function.

[0022] Furthermore, obtaining a disturbance sequence representing short-term rapid fluctuations of net load based on the trend sequence and the original net load sequence specifically includes:

[0023] Performing differential processing on the historical net load sequence and the trend sequence obtained by the regularization optimization method to obtain a disturbance sequence, which is the rapid fluctuation component of the original net load after removing the overall trend;

[0024] The calculation formula of the disturbance sequence is:

[0025] r(t)=x(t)-y(t)

[0026] Among them, x(t) represents the observed value of the historical net load series at the tth time point; y(t) represents the fitted value of the obtained trend series at the tth time point; r(t) represents the fitted value of the obtained disturbance series at the tth time point.

[0027] Furthermore, the construction of the trend prediction sub-network based on the series connection of the Transformer encoder and the long short-term memory network specifically includes:

[0028] The trend sequence obtained by the regularized optimization method is divided into multiple input samples according to the sliding window, and input into the trend prediction sub-network for training; the trend prediction sub-network is composed of a Transformer encoder and a long short-term memory network connected in sequence;

[0029] The Transformer encoder includes: a multi-head self-attention mechanism module for modeling the long-term dependencies between time steps in the trend sequence; a feedforward neural network module for performing nonlinear mapping on the features of each time step; a residual connection and layer normalization module for enhancing network stability and training deep structures; and a position encoding module for introducing temporal order information to compensate for the insensitivity of the Transformer structure to position information.

[0030] The Transformer encoder extracts features from the input trend sequence, and the output time series feature representation serves as the input of the subsequent LSTM network;

[0031] The long short-term memory network includes: a plurality of LSTM units, each unit including an input gate, a forget gate, and an output gate, for modeling and updating the time characteristics of the sequence; an output layer for mapping the LSTM output result of the final time step into the target trend prediction value;

[0032] The Transformer encoder and the LSTM network are sequentially connected through their feature output and input ends, so that the global dependency features extracted by the encoder are further modeled in the LSTM in terms of their local temporal dynamics, thereby improving the accuracy of trend prediction.

[0033] Furthermore, the trend sequence obtained from the historical net load sequence is used as a sample to train a trend prediction sub-network, specifically including:

[0034] The trend sequence obtained by decomposition using the regularized optimization method is used as the model input sample, the trend sequence corresponding to the future time step is used as the supervision label, the input-output pair is constructed for the trend sequence using a sliding window method, and the pair is input into the trend prediction sub-network for supervised training;

[0035] During the training process, the parameters of the trend prediction subnetwork are updated using the following loss function:

[0036]

[0037] Among them, L is the loss function of the trend prediction sub-network, N is the number of input-output pair samples constructed using the sliding window method, and y i 、 are the true trend value and predicted trend value of the i-th input-output pair sample, respectively. α and η>0 are loss weight adjustment parameters, which balance the contribution of the fitting error term and the smoothing term. β≥1 is the nonlinear power exponent of the fitting error, which is used to adjust the sensitivity to large errors and improve the robustness to outliers. γ>0 is the smooth coupling index, which is used to control the modulation effect of the change in the predicted trend at adjacent moments on the fitting error. θ≥1 is the nonlinear power exponent of the second-order difference smoothing of the trend, which strengthens the stationarity constraint of the trend series.

[0038] Furthermore, for the disturbance sequence, a regression prediction subnetwork based on historical environmental parameters is constructed, and the environmental parameters within the same historical window are used as input to train a disturbance prediction model, which specifically includes:

[0039] The environmental parameter sequence E before the historical moment t t ={e t-w+1 ,e t-w+2t ,…,e t} as input sample, where w is the length of the historical window, represents the environmental parameter vector at time t, and m is the environmental parameter dimension;

[0040] The disturbance sequence r corresponding to the historical moment t ={r t-w+1 ,r t-w+2 ,…,r t} as the regression target output, where r t is the disturbance value at time t;

[0041] Construct regression prediction subnetwork f θ (·), by mapping the environmental parameter sequence to the disturbance sequence, that is:

[0042]

[0043] in, represents the predicted disturbance value at time t+1, f θ represents the regression subnetwork;

[0044] The mean square error loss function is used to train the regression prediction subnetwork, and the loss function is expressed as:

[0045]

[0046] Among them, Lr is the loss function of the regression prediction sub-network, and M represents the number of training samples.

[0047] Furthermore, the regression prediction subnetwork includes a multi-layer fully connected neural network, each layer including a linear transformation, an activation function and a batch normalization module.

[0048] Furthermore, the trained trend prediction subnetwork and disturbance prediction model are used to respectively predict the trend component and disturbance component of each future time step and superimpose them to obtain the net load prediction value of multiple future time steps, specifically including:

[0049] Obtain the trend sequence data of the target power system at the current moment and before and the corresponding historical environmental parameters. The trend sequence data is recorded as y t-w+1:t =[y(t-w+1),y(t-w+2),…,y(t)], where y t-w+1:t Represents the trend sequence vector from time point t-w+1 to the current time t;

[0050] Construct an environmental parameter sequence E based on the environmental parameters t-w+1:t =[e(t-w+1),e(t-w+2),…,e(t)], where E t-w+1:t represents the environmental parameter sequence from time point t-w+1 to the current time t;

[0051] Using the trained trend prediction subnetwork, the trend component of the next k time steps is predicted based on the trend sequence input, which is expressed as:

[0052]

[0053] in, represents the forecast value sequence of the trend component; g θ represents the trend prediction subnetwork;

[0054] Using the trained disturbance prediction model, based on the constructed environmental parameter sequence input, the disturbance component of the next k time steps is predicted, which is expressed as:

[0055]

[0056] Among them, f θ represents the regression subnetwork; represents the predicted value sequence of the disturbance component;

[0057] The predicted trend component and disturbance component are superimposed point by point according to the time step to obtain the net load forecast value of the next k time steps, which is expressed as:

[0058]

[0059] in, represents the net load forecast sequence for the next k time steps.

[0060] Compared with the prior art, the present invention has the following advantages:

[0061] (1) Existing technologies usually use wavelet decomposition or empirical mode decomposition (such as EMD, EEMD, VMD, etc.) to perform multi-scale decomposition of power load series, dividing them into several high-frequency and low-frequency components to alleviate the non-stationarity of the data and improve the prediction accuracy. However, these methods have significant technical problems. First, such methods rely on a series of preset parameters, such as wavelet basis functions, number of decomposition layers or number of modes, and the selection of these parameters often lacks a unified standard and has a great impact on the final decomposition effect, making it difficult to stably adapt in different data scenarios. Second, the decomposition results of these methods lack clear physical interpretability. There is often a modal aliasing phenomenon between the trend term and the disturbance term, resulting in unclear boundaries between different frequency components. Short-term fluctuations or noise may remain in the trend part, affecting the performance of subsequent prediction models. In addition, existing methods find it difficult to balance the smoothness of the trend term with the fitting accuracy of the original sequence. It is easy for the trend curve to be too smooth and lose key change characteristics, or to be overly sensitive to abnormal points.

[0062] The present invention overcomes the above problems by constructing a regularized optimization objective function that takes into account both fitting accuracy and smoothness, thereby obtaining a trend sequence of the historical net load sequence. The method is centered on minimizing the residual between the trend term and the original net load sequence, while introducing a regularization term of the second-order difference of the trend sequence to constrain the smoothness of the trend, so that the solution can extract a smooth, continuous long-term trend with a small fitting error. The introduction of the Huber loss function further enhances the robustness of the model to mutation points and abnormal disturbances, avoiding the defect of the traditional square error being allergic to outliers. The regularized optimization model has a clear objective function structure and an adjustable smoothness control parameter. It not only has good mathematical solution properties, but also has feasibility and adaptability in engineering practice. Compared with wavelet decomposition and modal decomposition, which are highly empirical, complex to adjust parameters, and have unstable decomposition results, the trend extraction method of the present invention has higher stability, controllability and physical interpretability, laying a reliable data foundation for subsequent branch modeling and joint prediction of trends and disturbances, thereby significantly improving the accuracy and generalization ability of the overall prediction model.

[0063] (2) The present invention constructs a trend prediction sub-network based on the series connection of Transformer encoder and long short-term memory network (LSTM), aiming to solve the problem that traditional models are insufficient in modeling complex trend features. Existing methods such as using LSTM alone are prone to the problems of unstable long-term dependency modeling and inaccurate local change characterization. Although Transformer has global modeling capabilities, it is not sensitive to time sequence and has difficulty capturing fine-grained dynamic changes. The present invention first uses Transformer to extract the global dependency relationship in the trend sequence, and then uses LSTM to further model the local time series dynamics, taking into account the long-term changes and short-term fluctuation characteristics of the trend sequence, thereby improving the trend prediction accuracy and stability. In addition, position encoding strengthens the understanding of time sequence, and residual connection and layer normalization improve training efficiency and model depth expression ability. This series structure can effectively improve the multi-step prediction performance of the trend prediction sub-network for future trend components, providing more reliable trend support for the final net load prediction.

[0064] (3) The present invention aims to solve the problem that the disturbance component is greatly affected by external non-stationary environmental factors and is difficult to effectively model using traditional methods by constructing a regression prediction subnetwork based on historical environmental parameters for the disturbance sequence. The disturbance sequence reflects the rapid fluctuation of net load in the short term, and such fluctuations are often closely related to external environments such as temperature, humidity, wind speed, and solar irradiance. Existing methods usually regard disturbances as random noise or use a unified model for overall prediction, which cannot capture the nonlinear mapping relationship between disturbances and the environment, resulting in a decrease in prediction accuracy. The present invention constructs a regression prediction subnetwork with historical environmental parameters as input, so that the model can learn the implicit association between disturbances and environmental characteristics, and improve the ability to characterize the trend of disturbance changes. At the same time, the sliding window mechanism is used to construct input and output samples, so that the model has the ability to predict the short-term response to future disturbances, effectively enhancing the accuracy and stability of the overall net load prediction.

[0065] (4) The present invention decomposes the net load sequence into a trend sequence and a disturbance sequence, and uses networks with different structures for prediction respectively, in order to solve the problem that the net load sequence exhibits both trend and volatility, the dynamic change mechanism is complex, and a single model is difficult to take into account both long-term structure and short-term disturbance. The trend part reflects the steady change law of the net load evolution over time, which is suitable for modeling its long-term dependence through the Transformer encoder, and further capturing the short-term dynamics by LSTM to achieve high-precision modeling of trend evolution. The disturbance part reflects the short-term rapid changes caused by environmental factors, which has strong randomness and external drive. The present invention constructs a regression prediction subnetwork, takes historical environmental parameters as input, and specifically learns the sensitive response law of disturbance to external factors. Compared with the existing method of inputting the net load sequence as a whole into a single neural network model, which easily causes confusion between trend and disturbance and affects the modeling effect, the present invention not only improves the model's ability to express characteristics of different time scales through decomposition modeling and dual-branch prediction, but also significantly improves the accuracy and stability of multi-time step load prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 This is a flow chart of a net load prediction method according to an embodiment of the present invention;

[0067] Figure 2 This is a PMAE error frequency distribution diagram of the trend prediction subnetwork of an embodiment of the present invention and other models. DETAILED DESCRIPTION

[0068] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of the present invention.

[0069] Example 1:

[0070] This embodiment discloses a method for predicting net load of a power system based on regular decomposition and dual-branch prediction. Figure 1 As shown, the following steps are included:

[0071] Step S1: Obtain the target power system's historical net load sequence and corresponding historical environmental parameters. The historical net load sequence refers to a time series consisting of the target power system's actual observed net load values ​​within a certain time interval. This sequence reflects the system's net load changes at different time points and typically includes multiple consecutive sampling data points to describe the dynamic changes in the power system's load over a period of time. Historical environmental parameters include meteorological and renewable energy generation characteristic parameters related to net load changes, specifically temperature, humidity, wind speed, wind direction, and solar irradiance at the historical moment.

[0072] Step S2: Based on the regularized optimization method, an objective function is constructed that takes into account the smoothness of the fitting historical net load series and trend series. The trend series of the historical net load series is obtained by solving the objective function:

[0073] The objective function is:

[0074]

[0075] Where N is the total number of time points in the historical net load series; x(t) represents the observed value of the historical net load series at the tth time point; y(t) represents the fitted value of the trend series at the tth time point; λ>0 is the trend smoothing regularization coefficient; φ δ (r) is the Huber loss function, which is used to balance the sensitivity of the squared error to normal deviations and the robustness of the linear penalty to mutation points. It is defined as:

[0076]

[0077] Among them, r = x(t) - y(t) is the residual, and δ>0 is the residual threshold, which is used to distinguish normal deviations from abnormal disturbance mutations in the net load sequence.

[0078] The objective function is solved by a constrained numerical optimization algorithm, which is a gradient descent method, to obtain a trend sequence that minimizes the objective function.

[0079] Step S2 aims to address the existing problem of balancing fitting accuracy and smoothness in trend extraction by constructing an objective function based on a regularized optimization method that takes into account the smoothness of both the historical net load series and the trend series. Traditional methods such as wavelet decomposition or empirical mode decomposition can decompose the load series into different frequency components, but they often rely on complex parameter adjustments, and the trend part often has modal aliasing, resulting in a large amount of residual noise or short-term fluctuations in the trend series, affecting the accuracy and stability of the subsequent prediction model. In addition, existing methods have difficulty effectively handling outliers or mutations in the load series, which can easily distort trend estimates.

[0080] By setting the objective function, the present invention clearly balances the needs of two aspects: on the one hand, the Huber loss function is used to process the fitting residuals, which not only maintains the sensitivity to normal deviations but also enhances the robustness to abnormal mutation points, avoiding the excessive influence of abnormal disturbances on trend extraction; on the other hand, the second-order difference term of the trend sequence is introduced as a regularization term, and the smoothness of the trend sequence is constrained by adjusting the parameter λ, so that the trend change is more continuous and conforms to the long-term change law of the actual load, effectively avoiding overfitting of short-term fluctuations and causing the trend curve to be too uneven.

[0081] The objective function consists of two parts, corresponding to the fitting accuracy of the trend series and the smoothness of the trend, and the design takes these two key requirements into account. The first part is the measurement of the fitting error, which is in the form of The Huber loss function combines the advantages of square error and linear error. When the absolute value of the residual is less than the threshold δ, the loss is quadratic. Sensitive to normal deviations, which is conducive to accurately fitting most data points; when the residual exceeds δ, the loss is converted to a linear form Abnormally large residuals (such as mutation points or abnormal disturbances) are no longer amplified in a square form, but are linearly penalized, which enhances the robustness of the model to outliers and avoids the distortion of trend series due to extreme data points. This design enables trend extraction to accurately reflect the overall trend of most data and effectively suppress the interference of abnormal fluctuations on the trend. The second part is the trend smoothing regularization term This is the sum of the absolute values ​​of the second-order differences of the trend series. The second-order differences express the curvature of the trend series. This term penalizes the degree of curvature of the trend, promoting smoothness and continuity. The smoothness of the trend series can be controlled by adjusting the parameter λ: a larger λ enhances smoothness, resulting in a more gradual trend change and reducing the impact of short-term fluctuations. A smaller λ makes the trend series closer to the original data and preserves more detail. This regularization term helps eliminate noise and local fluctuations, ensuring that the trend series reflects the long-term evolution of the net load rather than short-term random disturbances.

[0082] Step S3: Obtain a disturbance sequence representing the short-term rapid fluctuation of the net load based on the trend sequence and the original net load sequence, specifically including:

[0083] The historical net load series is differentiated from the trend series obtained by the regularization optimization method to obtain the disturbance series, which is the rapid fluctuation component of the original net load after removing the overall trend.

[0084] The calculation formula of the disturbance sequence is:

[0085] r(t)=x(t)-y(t)

[0086] Among them, x(t) represents the observed value of the historical net load series at the tth time point; y(t) represents the fitted value of the obtained trend series at the tth time point; r(t) represents the fitted value of the obtained disturbance series at the tth time point.

[0087] Step S4: Construct a trend prediction sub-network based on the series connection of the Transformer encoder and the long short-term memory network, use the trend sequence obtained from the historical net load sequence as a sample, and train the trend prediction sub-network;

[0088] The trend sequence obtained by the regularized optimization method is divided into multiple input samples according to the sliding window and input into the trend prediction sub-network for training; the trend prediction sub-network is composed of a Transformer encoder and a long short-term memory network connected in sequence;

[0089] The Transformer encoder includes: a multi-head self-attention mechanism module for modeling the long-term dependencies between time steps in the trend sequence; a feedforward neural network module for nonlinear mapping of the features of each time step; a residual connection and layer normalization module for enhancing network stability and training deep structures; and a position encoding module for introducing temporal order information to compensate for the insensitivity of the Transformer structure to position information.

[0090] The Transformer encoder extracts features from the input trend sequence, and the output time series feature representation serves as the input of the subsequent LSTM network;

[0091] The long short-term memory network consists of several LSTM units, each of which contains an input gate, a forget gate, and an output gate, used to model the temporal characteristics of the sequence and update the memory; an output layer, used to map the LSTM output result of the final time step to the target trend prediction value;

[0092] The Transformer encoder and the LSTM network are sequentially connected through their feature output and input ends, so that the global dependency features extracted by the encoder can be further modeled in the LSTM to further model their local temporal dynamics, thereby improving the accuracy of trend prediction.

[0093] The trend sequence obtained by decomposition through regularized optimization method is used as the model input sample, and the trend sequence corresponding to the future time step is used as the supervision label. The input and output pairs of the trend sequence are constructed using the sliding window method and input into the trend prediction sub-network for supervised training;

[0094] During the training process, the parameters of the trend prediction subnetwork are updated using the following loss function:

[0095]

[0096] Among them, L is the loss function of the trend prediction sub-network, N is the number of input-output pair samples constructed using the sliding window method, and y i 、 are the true trend value and predicted trend value of the i-th input-output pair sample, respectively. α and η>0 are loss weight adjustment parameters, which balance the contribution of the fitting error term and the smoothing term. β≥1 is the nonlinear power exponent of the fitting error, which is used to adjust the sensitivity to large errors and improve the robustness to outliers. γ>0 is the smooth coupling index, which is used to control the modulation effect of the change in the predicted trend at adjacent moments on the fitting error. θ≥1 is the nonlinear power exponent of the second-order difference smoothing of the trend, which strengthens the stationarity constraint of the trend series.

[0097] Step S5: For the disturbance sequence, a regression prediction subnetwork based on historical environmental parameters is constructed. The environmental parameters within the same historical window are used as input to train a disturbance prediction model. The regression prediction subnetwork includes a multi-layer fully connected neural network, and each layer includes a linear transformation, an activation function, and a batch normalization module.

[0098] Step S5 specifically includes:

[0099] The environmental parameter sequence E before the historical moment t t ={e t-w+1 ,e t-w+2 ,...,e t} as input sample, where w is the length of the historical window, represents the environmental parameter vector at time t, and m is the environmental parameter dimension;

[0100] The disturbance sequence r corresponding to the historical moment t ={r t-w+1 ,r t-w+2 ,…,r t} as the regression target output, where r t is the disturbance value at time t;

[0101] Construct regression prediction subnetwork f θ (·), by mapping the environmental parameter sequence to the disturbance sequence, that is:

[0102]

[0103] in, represents the predicted disturbance value at time t+1, f θ represents the regression subnetwork;

[0104] The mean square error loss function is used to train the regression prediction subnetwork, and the loss function is expressed as:

[0105]

[0106] Among them, L r is the loss function of the regression prediction sub-network, and M represents the number of training samples.

[0107] Step S6: Using the trained trend prediction subnetwork and disturbance prediction model, the trend component and disturbance component of each future time step are predicted and superimposed to obtain the net load prediction value of multiple future time steps, which specifically includes:

[0108] Obtain the trend sequence data of the target power system at the current moment and before and the corresponding historical environmental parameters. The trend sequence data is recorded as y t-w+1:t =[y(t-w+1),y(t-w+2),...,y(t)], where y t-w+1:t Represents the trend sequence vector from time point t-w+1 to the current time t;

[0109] Construct environmental parameter sequence E based on environmental parameters t-w+1:t =[e(t-w+1),e(t-w+2),…,e(t)], where E t-w+1:t represents the environmental parameter sequence from time point t-w+1 to the current time t;

[0110] Using the trained trend prediction subnetwork, the trend component of the next k time steps is predicted based on the trend sequence input, which is expressed as:

[0111]

[0112] in, represents the forecast value sequence of the trend component; g θ represents the trend prediction subnetwork;

[0113] Using the trained disturbance prediction model, based on the constructed environmental parameter sequence input, the disturbance component of the next k time steps is predicted, which is expressed as:

[0114]

[0115] Among them, f θ represents the regression subnetwork; represents the predicted value sequence of the disturbance component;

[0116] The predicted trend component and disturbance component are superimposed point by point according to the time step to obtain the net load forecast value of the next k time steps, which is expressed as:

[0117]

[0118] in, represents the net load forecast sequence for the next k time steps.

[0119] This embodiment addresses the problem of the coexistence and complex changes of long-term trends and short-term disturbances in the net load forecast of the power system. It proposes a technical solution to decompose the net load sequence into a trend sequence and a disturbance sequence, and use different models for prediction. This decomposition and branch prediction approach overcomes the shortcomings of the existing technology that a single model directly predicts the overall net load sequence, which makes it difficult to take into account both long-term trends and short-term fluctuations. Traditional methods often ignore the non-stationary and multi-scale characteristics of net load data, resulting in limited model prediction results, and susceptibility to interference from abnormal points. The accuracy and stability are difficult to meet the needs of power grid scheduling.

[0120] This embodiment first extracts a smooth trend sequence that reflects long-term variations from historical net load data through an objective function based on regularized optimization. This objective function balances the accuracy of fitting historical net load data with the smoothness of the trend sequence, utilizes the Huber loss function to enhance robustness to abnormal disturbances, and simultaneously constrains the continuity and smoothness of trend variations through a second-order difference regularization term. This design addresses the technical issues of existing multi-scale decomposition methods, such as their reliance on a large number of empirical parameters, ambiguous decomposition results, and the inclusion of short-term fluctuations in trend terms. It achieves mathematical clarity and physical interpretability in the trend extraction process, ensures the stability and authenticity of the trend sequence, and provides a high-quality data foundation for subsequent trend prediction.

[0121] To predict trend sequences, this embodiment uses a trend prediction subnetwork that connects a Transformer encoder and a long short-term memory (LSTM) network in series. The Transformer's powerful global dependency modeling capabilities complement the LSTM's lack of sensitivity to temporal order, while the LSTM effectively captures local dynamic changes in trend sequences. This structure allows the model to not only accurately capture the global patterns of long-term trend changes, but also meticulously reflect short-term trend fluctuations, effectively improving the accuracy and stability of trend predictions and overcoming the limitations of traditional single models in capturing complex time series features.

[0122] For disturbance sequences, this embodiment constructs a regression prediction subnetwork based on historical environmental parameters, incorporating environmental factors such as weather and meteorology into disturbance prediction. This addresses the problem of traditional methods ignoring the impact of the external environment on short-term disturbances and treating disturbances as random noise, resulting in insufficient prediction accuracy. This subnetwork can learn the nonlinear mapping relationship between environmental parameters and disturbances, enhancing its responsiveness to short-term changes in disturbances and significantly improving the overall net load forecast's ability to capture details and accuracy.

[0123] This embodiment effectively addresses the modeling challenges posed by the coexistence of long-term stability and short-term non-stationarity in net load series through the decomposition of trends and disturbances and a dual-branch network forecasting strategy. While improving forecast accuracy, this solution enhances the model's robustness to abnormal fluctuations and environmental disturbances, and improves the stability and generalization capabilities of multi-time-step forecasts. Overall, this embodiment not only optimizes the data processing process and enhances the model's expressiveness, but also achieves more scientific and reasonable load forecasting, promoting technological progress in intelligent power system scheduling and renewable energy consumption.

[0124] Example 2:

[0125] The present invention proposes a trend prediction subnetwork (LTE model) based on the connection of Transformer encoder and long short-term memory network in series. This novel architecture aims to efficiently capture and utilize the time dependency in multivariate time series data to enhance the accuracy of trend series. This study obtains relevant electricity consumption data through smart grids. In the experiment, public historical electricity load consumption data was used. It includes various building types (residential and commercial) and their energy consumption time series data in different climatic regions of my country. In the study, the optimal lookback window for each implementation model was determined by data obtained from 650 groups of data pools. The size of each group of data is different, which is achieved by random selection of the entire data pool. Then, each prediction model in all clusters is trained and evaluated.

[0126] Table 1. PMAE values ​​of the proposed LTE and other prediction models. Average optimal input sequence lengths under these different settings. The performance of each model is evaluated both per state and overall, focusing on achieving good accuracy on most of the data while avoiding extensive individual optimization.

[0127] The optimal lookback window for each implementation model was determined using data obtained from the smart grid. Each prediction model was then trained using these data clusters. Table 1 presents a comparative analysis of the performance of various models.

[0128] Table 1 PMAE values ​​of the proposed LTE and other prediction models

[0129]

[0130] Table 1 presents a comparative analysis of the performance of various models, showing PMAE values ​​for the proposed LTE and other prediction models. These models, including those listed above, were evaluated on combinations of the three data clusters using the defined percentage MAE metric. Models are ranked by their overall PMAE performance. Columns 2, 3, and 4 report each model's PMAE performance for the three data clusters, while the fifth column reports the combined error for these three data clusters. The last column reports the standard deviation of the prediction error distribution across the entire dataset. The LTE and LT models consistently outperformed the other models, demonstrating superior performance. Their overall PMAE error was 1. However, the training and inference times for LT increased by an average of 28.7% compared to LTE. Overall, the enhanced LSTM structure failed to significantly impact the results when used as the initial component of the hybrid LT architecture, regardless of whether it was paired with the LT or LTE variants.

[0131] Figure 2 In this paper, 150 data samples of different sizes were predicted using different models to compare the differences between different models. The results are as follows: Figure 2 As shown. Figure 2 All models and experiments are covered, so it represents a large and diverse dataset, highlighting the wide applicability and reliability of LTE models under different conditions. Figure 2 Subfigures (b)-(f) in Figure 3 show the percentage frequency distribution of prediction errors of models ranked lower than LTE. Figure 2 Subfigures (g)-(k) show the difference in error between the LTE model and the next top five models, quantified in units of PMAE × 10^-5. These graphs show the proportion of times the LTE model's error is lower (for positive examples) or higher (for negative examples) than the comparison model. This difference serves as an indicator of how often the LTE model's predictions are closer to the ground truth.

[0132] If the above functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0133] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and such modifications or substitutions are intended to be within the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be subject to the scope of protection of the claims.

Claims

1. A method for predicting net load of a power system based on regularized decomposition and dual-branch prediction, characterized in that: The following steps are involved: Obtain the historical net load sequence and corresponding historical environmental parameters of the target power system; Based on the regularized optimization method, an objective function is constructed that takes into account the smoothness of the fitting of the historical net load series and the trend series. The trend series of the historical net load series is obtained by solving the objective function. Obtaining a disturbance sequence representing short-term rapid fluctuations of the net load according to the trend sequence and the original net load sequence; A trend prediction subnetwork based on a series connection between a Transformer encoder and a long short-term memory network is constructed. The trend sequence obtained from the historical net load sequence is used as a sample to train the trend prediction subnetwork. For the disturbance sequence, a regression prediction subnetwork based on historical environmental parameters is constructed, and the environmental parameters within the same historical window are used as input to train a disturbance prediction model; The trained trend prediction subnetwork and disturbance prediction model are used to predict the trend component and disturbance component of each future time step respectively and superimpose them to obtain the net load prediction value of multiple future time steps.

2. A method for predicting net load of a power system based on regularized decomposition and dual-branch prediction according to claim 1, characterized in that: The historical environmental parameters include meteorological and renewable energy power generation characteristic parameters related to net load changes, specifically including: temperature, humidity, wind speed, wind direction, and solar irradiance at historical moments.

3. The method for predicting net load of a power system based on regularized decomposition and dual-branch prediction according to claim 1, characterized in that: The objective function of constructing the smoothness of the historical net load series and trend series is: Where N is the total number of time points in the historical net load series; x(t) represents the observed value of the historical net load series at the tth time point; y(t) represents the fitted value of the trend series at the tth time point; λ>0 is the trend smoothing regularization coefficient; φ δ (r) is the Huber loss function, which is used to balance the sensitivity of the squared error to normal deviations and the robustness of the linear penalty to mutation points. It is defined as: Among them, r = x(t) - y(t) is the residual, and δ>0 is the residual threshold, which is used to distinguish normal deviations from abnormal disturbance mutations in the net load sequence.

4. The method for predicting net load of a power system based on regularized decomposition and dual-branch prediction according to claim 1, characterized in that: The method of obtaining the trend sequence of the historical net load sequence by solving the objective function specifically includes: solving the objective function using a constrained numerical optimization algorithm, wherein the numerical optimization algorithm is a gradient descent method, to obtain the trend sequence that minimizes the objective function.

5. The method for predicting net load of a power system based on regularized decomposition and dual-branch prediction according to claim 1, characterized in that: Obtaining a disturbance sequence representing short-term rapid fluctuations of the net load based on the trend sequence and the original net load sequence specifically includes: Performing differential processing on the historical net load sequence and the trend sequence obtained by the regularization optimization method to obtain a disturbance sequence, which is the rapid fluctuation component of the original net load after removing the overall trend; The calculation formula of the disturbance sequence is: r(t)=x(t)-y(t) Among them, x(t) represents the observed value of the historical net load series at the tth time point; y(t) represents the fitted value of the obtained trend series at the tth time point; r(t) represents the fitted value of the obtained disturbance series at the tth time point.

6. The method for predicting net load of a power system based on regularized decomposition and dual-branch prediction according to claim 1, characterized in that: The construction of the trend prediction sub-network based on the series connection of Transformer encoder and long short-term memory network specifically includes: The trend sequence obtained by the regularized optimization method is divided into multiple input samples according to the sliding window, and input into the trend prediction sub-network for training; the trend prediction sub-network is composed of a Transformer encoder and a long short-term memory network connected in sequence; The Transformer encoder includes: a multi-head self-attention mechanism module for modeling the long-term dependencies between time steps in the trend sequence; a feedforward neural network module for performing nonlinear mapping on the features of each time step; a residual connection and layer normalization module for enhancing network stability and training deep structures; and a position encoding module for introducing temporal order information to compensate for the insensitivity of the Transformer structure to position information. The Transformer encoder extracts features from the input trend sequence, and the output time series feature representation serves as the input of the subsequent LSTM network; The long short-term memory network includes: a plurality of LSTM units, each unit including an input gate, a forget gate, and an output gate, for modeling and updating the time characteristics of the sequence; an output layer for mapping the LSTM output result of the final time step into the target trend prediction value; The Transformer encoder and the LSTM network are sequentially connected through their feature output and input ends, so that the global dependency features extracted by the encoder are further modeled in the LSTM in terms of their local temporal dynamics, thereby improving the accuracy of trend prediction.

7. The method for predicting net load of a power system based on regularized decomposition and dual-branch prediction according to claim 1, characterized in that: The trend sequence obtained from the historical net load sequence is used as a sample to train a trend prediction sub-network, specifically including: The trend sequence obtained by decomposition using the regularized optimization method is used as the model input sample, the trend sequence corresponding to the future time step is used as the supervision label, the input-output pair is constructed for the trend sequence using a sliding window method, and the pair is input into the trend prediction sub-network for supervised training; During the training process, the parameters of the trend prediction subnetwork are updated using the following loss function: Among them, L is the loss function of the trend prediction sub-network, N is the number of input-output pair samples constructed using the sliding window method, and y i 、 are the true trend value and predicted trend value of the i-th input-output pair sample, respectively. α and η>0 are loss weight adjustment parameters, which balance the contribution of the fitting error term and the smoothing term. β≥1 is the nonlinear power exponent of the fitting error, which is used to adjust the sensitivity to large errors and improve the robustness to outliers. γ>0 is the smooth coupling index, which is used to control the modulation effect of the change in the predicted trend at adjacent moments on the fitting error. θ≥1 is the nonlinear power exponent of the second-order difference smoothing of the trend, which strengthens the stationarity constraint of the trend series.

8. The method for predicting net load of a power system based on regularized decomposition and dual-branch prediction according to claim 1, characterized in that: The method of constructing a regression prediction subnetwork based on historical environmental parameters for the disturbance sequence, taking the environmental parameters in the same historical window as input, and training a disturbance prediction model specifically includes: The environmental parameter sequence E before the historical moment t t ={e t-w+1 ,e t-w+2 ,...,e t } as input sample, where w is the length of the historical window, represents the environmental parameter vector at time t, and m is the environmental parameter dimension; The disturbance sequence r corresponding to the historical moment t ={r t-w+1 ,r t-w+2 ,…,r t } as the regression target output, where r t is the disturbance value at time t; Construct regression prediction subnetwork f θ (·), by mapping the environmental parameter sequence to the disturbance sequence, that is: in, represents the predicted disturbance value at time t+1, f θ represents the regression subnetwork; The mean square error loss function is used to train the regression prediction subnetwork, and the loss function is expressed as: Among them, L r is the loss function of the regression prediction sub-network, and M represents the number of training samples.

9. The method for predicting net load of a power system based on regularized decomposition and dual-branch prediction according to claim 1, characterized in that: The regression prediction subnetwork includes a multi-layer fully connected neural network, each layer including a linear transformation, an activation function and a batch normalization module.

10. The method for predicting net load of a power system based on regularized decomposition and dual-branch prediction according to claim 1, characterized in that: The trained trend prediction subnetwork and disturbance prediction model are used to predict the trend component and disturbance component of each future time step respectively and superimpose them to obtain the net load prediction value of multiple future time steps, specifically including: Obtain the trend sequence data of the target power system at the current moment and before and the corresponding historical environmental parameters. The trend sequence data is recorded as y t-w+1:t =[y(t-w+1),y(t-w+2),...,y(t)], where y t-w+1:t Represents the trend sequence vector from time point t-w+1 to the current time t; Construct an environmental parameter sequence E based on the environmental parameters t-w+1:t =[e(t-w+1),e(t-w+2),…,e(t)], where E t-w+1:t represents the environmental parameter sequence from time point t-w+1 to the current time t; Using the trained trend prediction subnetwork, the trend component of the next k time steps is predicted based on the trend sequence input, which is expressed as: in, represents the forecast value sequence of the trend component; g θ represents the trend prediction subnetwork; Using the trained disturbance prediction model, based on the constructed environmental parameter sequence input, the disturbance component of the next k time steps is predicted, which is expressed as: Among them, f θ represents the regression subnetwork; represents the predicted value sequence of the disturbance component; The predicted trend component and disturbance component are superimposed point by point according to the time step to obtain the net load forecast value of the next k time steps, which is expressed as: in, represents the net load forecast sequence for the next k time steps.

Images