A power energy supply and dispatch prediction method

Abstract

This application provides a method for predicting power energy supply and dispatch, belonging to the field of time series forecasting technology. The method includes: acquiring historical power generation parameters and dividing them into segments encoded as power time-slice sequences; performing anti-disturbance processing on the power time-slice sequences, and performing feature incrementing and bit-by-bit modulation to obtain bit-by-bit modulated features; capturing the time embedding vector of the power time-slice sequences through a time-modulated neural network, and generating an output response by combining basic linear terms and spline terms; generating grid nodes through an adaptive grid update mechanism, and generating a uniform grid node sequence through a uniform grid; linearly weighting and fusing the grid nodes and uniform grid node sequences to obtain a fused grid node sequence; performing boundary expansion and least-squares refitting of the spline weights on the fused grid node sequence to obtain an expanded node sequence, and updating the spline function weights to generate a prediction result representing future power generation. This improves the accuracy of power energy forecasting.

Description

Technical Field

[0001] This invention belongs to the field of time series forecasting technology, and in particular relates to a method for forecasting power energy supply and dispatch. Background Technology

[0002] With the accelerated advancement of the intelligent transformation of the power system, accurate prediction of power generation has become a key technology for ensuring the stable operation of the power grid and optimizing energy dispatch. The large-scale grid connection of new energy sources such as wind power and photovoltaics has resulted in strong fluctuations and intermittent nature in power generation output, posing a severe challenge to traditional prediction methods.

[0003] In existing technologies, data anomalies and signal loss can lead to significant deviations in prediction results.

[0004] Therefore, the present invention provides a method for predicting power energy supply and dispatch. Summary of the Invention

[0005] This invention provides a method for predicting power energy supply and dispatch, which at least solves the problem of inaccurate prediction results in the prior art.

[0006] This application provides a method for predicting power energy supply and dispatch, the method comprising: Step S1: Obtain historical power generation parameters, divide each historical power generation parameter into segments and encode them into power time slice sequences; Step S2: Use a soft forgetting adapter to perform anti-disturbance processing on the power time slice sequence to obtain the perturbed features, and then perform feature increment and bit-by-bit modulation on the perturbed features in sequence to obtain the bit-by-bit modulated features. Step S3: Capture the temporal embedding vector of the power time slice sequence using a time-modulated neural network, and generate the output response by combining the basic linear term and spline term; Step S4: Generate grid nodes through an adaptive grid update mechanism, and generate a uniform grid node sequence through a uniform grid. Then, perform linear weighted fusion of the grid nodes and the uniform grid node sequence to obtain a fused grid node sequence. Step S5: Perform boundary expansion and least-squares refitting of spline weights on the fused grid node sequence to obtain the expanded node sequence, and update the spline function weights to generate prediction results for characterizing future power generation.

[0007] Furthermore, the soft forgetting adapter described in step S2 perturbs the power time-slice sequence by generating a forgetting mask: The forgetting mask is a mask generated by point forgetting mode or block forgetting mode. In point forgetting mode, each time step is sampled independently, and the mask elements are generated by Bernoulli distribution. The formula for generating the point forgetting mode mask is:

[0008] in, The sample number. For time step index, For batch size, For time step; The block forgetting pattern forgets a random starting position for each sample, with a length of [length missing]. For continuous intervals, the formula for generating the block forgetting pattern mask is:

[0009]

[0010] in, For the first The starting position of the forgotten segment in each sample.

[0011] Furthermore, in step S2, a soft forgetting adapter is used to perform anti-disturbance processing on the power time-slice sequence to obtain the perturbed features, the expression of which is:

[0012] in, The characteristics after the disturbance; It is a power time slice sequence; Forgetting intensity, ; The expression for the forget mask is: .

[0013] Further, in step S2, the perturbed features are sequentially subjected to feature increment and bit-by-bit modulation to obtain the bit-by-bit modulated features, specifically including: Step S21: Map the perturbed features through the bottleneck adapter path to obtain the feature increment term, the calculation formula of which is:

[0014] in, Represents the feature increment term. Represents a dimension-reduced linear mapping. Represents a linear mapping of increasing dimensions. It is a non-linear activation function; Step S22: Modulate the feature increment term and the forget mask by element-wise multiplication to obtain the modulated increment term; then add the modulated increment term to the perturbed feature to obtain the bit-modulated feature.

[0015] Further, in step S22, the feature increment term and the forgetting mask are modulated by element-wise multiplication to obtain the modulated increment term; then, the modulated increment term is added to the perturbed feature to obtain the bit-modulated feature, the expression of which is:

[0016] in, This represents the bit-modulated features, used to characterize more robust intermediate timing information under complex interference scenarios, and serves as the input to the time-modulated neural network in step S3 to generate the output response.

[0017] Further, in step S3, the time embedding vector of the power time slice sequence is captured by a time modulation neural network, the expression of which is:

[0018] in, It is a time step Temporal embedding vector; Indicates from time index The embedding vector obtained from the lookup table is mapped to ; It is a linear mapping matrix. .

[0019] Furthermore, in step S3, the output response is generated by combining the basic linear term and the spline term:

[0020]

[0021]

[0022] in, The output response characterizes the output response of the power time-slice sequence at the current time step after the combined effects of the basic linear term, spline term, and time modulation. Basic linear terms; For spline terms; The learnable parameter matrix of the basic linear terms; The learnable parameter matrix for the spline terms. These are basic activation functions such as SiLU; This represents the expansion of the spline basis functions for each input dimension.

[0023] Further, in step S4, the time-modulated neural network includes an adaptive grid update mechanism, which generates grid nodes, and its expression is:

[0024] in, This is an adaptive grid node sequence used to adaptively determine the grid node positions based on the actual distribution of the input features. For the number of grid segments, For quantile sampling index, For the first After sorting the dimensional features, in the position The quantile values ​​obtained at the location.

[0025] Furthermore, a uniform grid node sequence is generated using a uniform grid, the expression of which is:

[0026]

[0027] in, A uniform grid node sequence, used to provide reference grid nodes evenly distributed within the boundary interval. It is an equally spaced grid index; and These are all boundary values, and their expressions are:

[0028] in, This is the boundary expansion amount, used to expand the value range of the input feature when constructing a uniform grid node sequence. For the first The maximum value after sorting the features. For the first The minimum value after sorting the features.

[0029] Further, in step S4, the grid node sequence and the uniform grid node sequence are linearly weighted and fused to obtain the fused grid node sequence, the expression of which is:

[0030] in, To merge the grid node sequences, This is the mesh smoothing fusion factor.

[0031] As can be seen from the above technical solutions, the present invention has the following advantages: The power energy supply and dispatch prediction method provided in this application uses a soft forgetting adapter to perform anti-disturbance processing on the power time slice series, which effectively improves the accuracy of future power energy prediction under unreliable scenarios such as missing data and abnormal input. The adapter simulates local data loss caused by sensor failure or communication interruption through a forgetting mask mechanism, and can learn robust feature representations under partial observation conditions, thereby significantly reducing the impact of noise interference on the prediction results and ensuring stable output in real power data acquisition environment.

[0032] By capturing time-embedded vectors through time-modulated neural networks and combining them with spline basis functions, the dynamic modeling capability for non-stationary time-series characteristics of power generation is enhanced. The nonlinear mapping relationship of different time periods is adaptively adjusted through time-aware spline kernel functions, which accurately captures the power change patterns caused by external factors such as sudden weather changes and load fluctuations, overcoming the shortcomings of traditional fixed-parameter models in fitting long-term trend drift and short-term sudden patterns.

[0033] The adaptive grid update mechanism optimizes the accuracy and stability of local feature fitting by fusing data-driven grids and uniform grids. It dynamically adjusts the grid node density according to the distribution of input power values, maintains computational efficiency in the power flat range, and enhances sensitivity in the range of drastic fluctuations. This improves the tracking ability of extreme values ​​while ensuring global approximation accuracy and reducing prediction bias. Attached Figure Description

[0034] To more clearly illustrate the technical solution of this application, the accompanying drawings used in the description will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0035] Figure 1 This is a schematic diagram illustrating the framework principle of the power energy supply and dispatch prediction method described in this invention.

[0036] Figure 2 This is a structural diagram of the soft forgetting adapter in the power energy supply and dispatch prediction method of the present invention.

[0037] Figure 3 This is a diagram of the time-modulated neural network structure of the power energy supply and dispatch prediction method described in this invention. Detailed Implementation

[0038] To make the purpose, features, and advantages of this application more apparent and understandable, specific embodiments and accompanying drawings will be used to clearly and completely describe the technical solution protected by this application. Obviously, the embodiments described below are only some embodiments of this application, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0039] This application provides a method for predicting power energy supply and dispatch, addressing the urgent technical problem of accurately predicting power energy.

[0040] The technical solutions proposed in the embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0041] This application provides a method for predicting power energy supply and dispatch, which specifically includes the following steps: Step S1: Obtain historical power generation parameters, divide each historical power generation parameter into segments and encode them into power time slice sequences; Step S2: The power time slice sequence is subjected to anti-disturbance processing using a soft forgetting adapter to obtain the perturbed features. The perturbed features are then subjected to feature increment and bit-by-bit modulation in sequence to obtain the bit-by-bit modulated features. The bit-by-bit modulated features are used to characterize the changing trend of power generation at the multiple historical times.

[0042] Step S3: Capture the temporal embedding vector of the power time slice sequence using a time-modulated neural network, and generate the output response by combining the basic linear term and spline term; Step S4: Generate grid nodes through an adaptive grid update mechanism, and generate a uniform grid node sequence through a uniform grid. Then, perform linear weighted fusion of the grid nodes and the uniform grid node sequence to obtain a fused grid node sequence. Step S5: Perform boundary expansion and least-squares refitting of spline weights on the fused grid node sequence to obtain the expanded node sequence, and update the spline function weights to generate prediction results for characterizing future power generation.

[0043] In this embodiment, a power time slice sequence is given. ,in It is the batch size. For time step, This refers to the time slice dimension. First, the time series of power generation is divided into slices of length [length missing]. The patch fragment is encoded as ,in , This represents the amount of power generated. Based on this, the model updates the time slice representation layer by layer and outputs predictions to characterize future power generation.

[0044] like Figure 1 As shown, this invention divides the input power operation parameters into equal-length patch segments, which are then sequentially fed into a backbone network composed of multiple temporal layers. Each temporal layer consists of three core components: a TimeAttention mechanism, a Soft Forget Adapter (SFA), and a Temporally Modulated Neural Network (TMKAN). The TimeAttention mechanism captures long-range dependencies in the time series; the SFA actively simulates unreliable input scenarios during the training phase through random masking and residual modulation mechanisms, guiding the model to learn robust representations under incomplete context conditions; and the TMKAN is embedded in the feedforward structure of each Transformer layer, using a time-aware spline kernel function to achieve nonlinear time-varying mapping between input variables. The three modules work synergistically to improve the accuracy of power generation prediction data. By simulating missing data in the time series through the SFA and structurally perturbing intermediate features, the model is guided to learn structural reasoning capabilities for masked regions, thereby improving prediction performance under unreliable inputs.

[0045] like Figure 2 As shown, the soft forgetting adapter (SFA) consists of two key parts: a controllable masking generator and a bottleneck adapter path. Specifically, it first accepts the power time slice sequence input by TimeAttention. In the controlled masking generation stage, the soft forgetting adapter (SFA) provides two forgetting modes: a point forgetting mode where forgetting is generated by independently sampling at each time step, and a block forgetting mode where forgetting occurs over a randomly selected fixed-length time period. The soft forgetting adapter (SFA) adjusts the forgetting based on the forgetting ratio. Control the sample forgetting rate, and according to Controlling the intensity of forgetting. The soft forgetting adapter described in step S2 perturbs the power time-slice sequence by generating a forgetting mask: The forgetting mask is a mask generated by either point forgetting mode or block forgetting mode. Both modes are used in constructing the forgetting mask. The differences are as follows: specifically, given a forgetting rate... Then, the point-forgetting mode samples independently at each time step, and the mask's first... Whether an element is masked is generated by a Bernoulli distribution.

[0046] The formula for generating the point forgetting pattern mask is:

[0047] in, The sample number. For time step index, For batch size, For time step; The formula for generating the block forgetting pattern mask is:

[0048]

[0049] in, For the first The starting position of the forgotten segment in each sample; In the point-forgetting mode, each time step is sampled independently, and the mask elements are generated by a Bernoulli distribution; in the block-forgetting mode, each sample is forgotten at a random starting position with a length of [missing information]. Continuous intervals.

[0050] In step S2, a soft forgetting adapter is used to perform anti-disturbance processing on the power time-slice sequence to obtain the perturbed features:

[0051] in, The characteristics after the disturbance; It is a power time slice sequence; Forgetting intensity, ;when When, it degenerates into the original input; when When, it relies entirely on the mask generated by the forgetting mode to achieve strong perturbation; when When forgotten, the location will be retained. Proportional information allows for soft forgetting.

[0052] The expression for the forget mask is: .

[0053] In step S2, the perturbed features are sequentially subjected to feature increment and bit-by-bit modulation to obtain the bit-by-bit modulated features, specifically including: Step S21: Map the perturbed features through the bottleneck adapter path to obtain the feature increment term, the calculation formula of which is:

[0054] in, Represents the feature increment term. Represents a dimension-reduced linear mapping. Represents a linear mapping of increasing dimensions. It is a non-linear activation function; Step S22: Modulate the feature increment term and the forget mask by element-wise multiplication to obtain the modulated increment term; then add the modulated increment term to the perturbed feature to obtain the bit-modulated feature.

[0055] In step S22, the feature increment term and the forgetting mask are modulated by element-wise multiplication to obtain the modulated increment term; then, the modulated increment term is added to the perturbed feature to obtain the bit-modulated feature, the expression of which is:

[0056] in, This represents the bit-modulated features, used to characterize more robust intermediate timing information under complex interference scenarios, and serves as the input to the time-modulated neural network in step S3 to generate the output response.

[0057] In this invention, the soft forgetting adapter (SFA) and the time-modulated neural network (TMKAN) work together. The former enhances the model's ability to resist disturbances from unreliable inputs, while the latter enhances its ability to express time-series non-stationary changes, thus jointly improving the prediction performance of power operation parameters.

[0058] To enhance the model's ability to represent non-stationary changes in time-series data, this invention proposes a Temporal Modulation Neural Network (TMKAN) based on the Kolmogorov–Arnold theorem. This TMKAN is embedded into the feedforward structure of a Transformer, replacing the traditional MLP structure, to enhance the model's adaptability and generalization ability in dynamic sequence modeling.

[0059] like Figure 3 As shown, the Time-Modulated Neural Network (TMKAN) mainly consists of two parts: a time modulation mechanism and an adaptive grid update mechanism. The time modulation mechanism maps the time index to a time modulation vector through lookup table embedding and a small projection network, and adds it to the spline output in the form of residual modulation. This is used to adjust the nonlinear response strength of each connection, allowing the activation function to adapt to time. The adaptive grid update mechanism, while retaining the structure of the TMKAN that explicitly models a univariate nonlinear function for each input dimension, adaptively constructs a new grid update mechanism based on the distribution characteristics of the current input. The spline grid nodes are refitted and the spline weights are refitted, thereby enhancing the model's expressive flexibility in local regions and its dynamic adaptability to distribution changes.

[0060] Specifically, the Temporal Modulation Neural Network (TMKAN) introduces a temporal embedding vector while maintaining the basic structure of the KAN. Dynamically modulate the activation mapping, assuming the input sample is... Its corresponding time step is In step S3, the time embedding vector of the power time slice sequence is captured by a time modulation neural network, and its expression is:

[0061] in, It is a time step The time embedding vector is used to control the output response; Indicates from time index The embedding vector obtained from the lookup table is mapped to ; It is a linear mapping matrix. .

[0062] The time embedding vector is added as a global conditional residual term in each KAN layer to the sum of the basic linear term and the spline term to form the output response. In step S3, the output response is generated by combining the basic linear term and the spline term.

[0063]

[0064]

[0065] in, The output response is used to characterize the output response of the power time slice sequence at the current time step after the combined effects of the basic linear term, spline term, and time modulation, thereby improving the ability of the spline term to characterize local regions with dense features. Basic linear terms; For spline terms; The learnable parameter matrix of the basic linear terms; The learnable parameter matrix for the spline terms. These are basic activation functions such as SiLU; The structure represents the expansion of the spline basis function for each input dimension. This structure can be viewed as a channel-level residual modulation, introducing time context adaptability with minimal modeling cost.

[0066] To enhance the model's local expressive power under non-stationary changes, this embodiment introduces an adaptive grid update mechanism. During each layer of feedforward, the spline node positions are adaptively adjusted based on the power time-slice sequence distribution, thereby enhancing the flexibility of local function approximation and the model's generalization ability. For each dimension of the power time-slice sequence... First, for the input in the current batch Sort in ascending order and sample Each quantile constitutes a node sequence. In step S4, the time-modulated neural network includes an adaptive grid update mechanism, which generates grid nodes. The expression for this mechanism is:

[0067] in, This is an adaptive grid node sequence used to adaptively determine the grid node positions based on the actual distribution of the input features. For the number of grid segments, For quantile sampling index, For the first After sorting the dimensional features, in the position The quantile values ​​obtained at the location.

[0068] The expression for generating a uniform grid node sequence using a uniform grid is:

[0069]

[0070] in, A uniform grid node sequence is used to provide reference grid nodes that are evenly distributed within the boundary intervals, ensuring the overall stability and regularity of the mesh generation. It is an equally spaced grid index; and These are all boundary values, and their expressions are:

[0071] in, This is the boundary expansion amount, used to expand the value range of the input feature when constructing a uniform grid node sequence. For the first The maximum value after sorting the features. For the first The minimum value after sorting the features.

[0072] In step S4, the grid node sequence and the uniform grid node sequence are linearly weighted and fused to obtain the fused grid node sequence, the expression of which is:

[0073] in, To merge the grid node sequences, This is the mesh smoothing fusion factor.

[0074] Finally, the fused mesh node sequence is subjected to boundary expansion and least-squares refitting with spline weights. At order [order missing] In the case of expansion, we obtain the expanded node sequence. And reconstruct the spline function weights accordingly. .

[0075] Spline weights are updated by least-squares fitting of the original function curve, with the objective function being:

[0076] in, This represents the output of the unreduced spline calculated under the old grid. For the first The basis functions are at position The value on, These are the corresponding weighting coefficients.

[0077] The final output is the power generation forecast result for the future forecast period. In power energy forecasting, the power generation forecast result is input to the dispatch decision module to support dispatch operations such as load allocation, grid connection dispatch and reserve capacity configuration, thereby reducing the risk of supply and demand imbalance and improving the dispatch stability and operational reliability of the power system in complex operating environments.

[0078] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

[0079] Any changes, modifications, substitutions, and variations made to the embodiments without departing from the principles and spirit of the present invention still fall within the protection scope of the present invention.

Claims

1. A method for power supply and dispatch forecasting, characterized in that, The method includes: Step S1: Obtain historical power generation parameters, divide each historical power generation parameter into segments and encode them into power time slice sequences; Step S2: Use a soft forgetting adapter to perform anti-disturbance processing on the power time slice sequence to obtain the perturbed features, and then perform feature increment and bit-by-bit modulation on the perturbed features in sequence to obtain the bit-by-bit modulated features. Step S3: Capture the temporal embedding vector of the power time slice sequence using a time-modulated neural network, and generate the output response by combining the basic linear term and spline term; Step S4: Generate grid nodes through an adaptive grid update mechanism, and generate a uniform grid node sequence through a uniform grid. Then, perform linear weighted fusion of the grid nodes and the uniform grid node sequence to obtain a fused grid node sequence. Step S5: Perform boundary expansion and least-squares refitting of spline weights on the fused grid node sequence to obtain the expanded node sequence, update the spline function weights, and generate prediction results to characterize future power generation.

2. The method of claim 1, wherein, The soft forgetting adapter described in step S2 perturbs the power time-slice sequence by generating a forgetting mask: The forgetting mask is a mask generated by point forgetting mode or block forgetting mode. In point forgetting mode, each time step is sampled independently, and the mask elements are generated by Bernoulli distribution. The formula for generating the point forgetting mode mask is: wherein, is a sample number, is a time step index, is a batch size, is a time step length; The block forgetting pattern forgets a random starting position for each sample, with a length of [length missing]. For continuous intervals, the formula for generating the block forgetting pattern mask is: in, For the first The starting position of the forgotten segment in each sample.

3. The power energy supply and dispatch forecasting method as described in claim 2, characterized in that, In step S2, a soft forgetting adapter is used to perform anti-disturbance processing on the power time-slice sequence to obtain the perturbed features, the expression of which is: in, The characteristics after the disturbance; It is a power time slice sequence; Forgetting intensity, ; The expression for the forget mask is: 。 4. The power energy supply and dispatch forecasting method as described in claim 3, characterized in that, In step S2, the perturbed features are sequentially subjected to feature increment and bit-by-bit modulation to obtain the bit-by-bit modulated features, specifically including: Step S21: Map the perturbed features through the bottleneck adapter path to obtain the feature increment term, the calculation formula of which is: in, Represents the feature increment term. Represents a dimension-reduced linear mapping. Represents a linear mapping of increasing dimensions. It is a non-linear activation function; Step S22: Modulate the feature increment term and the forget mask by element-wise multiplication to obtain the modulated increment term; then add the modulated increment term to the perturbed feature to obtain the bit-modulated feature.

5. The power energy supply and dispatch forecasting method as described in claim 4, characterized in that, In step S22, the feature increment term and the forgetting mask are modulated by element-wise multiplication to obtain the modulated increment term; then, the modulated increment term is added to the perturbed feature to obtain the bit-modulated feature, the expression of which is: in, This represents the bit-modulated features, used to characterize more robust intermediate timing information under complex interference scenarios, and serves as the input to the time-modulated neural network in step S3 to generate the output response.

6. The power energy supply and dispatch forecasting method as described in claim 5, characterized in that, In step S3, the time embedding vector of the power time slice sequence is captured by a time modulation neural network, and its expression is: in, It is a time step Temporal embedding vector; Indicates from time index The embedding vector obtained from the lookup table is mapped to ; It is a linear mapping matrix. .

7. The power energy supply and dispatch forecasting method as described in claim 6, characterized in that, In step S3, the output response is generated by combining the basic linear term and the spline term: in, The output response characterizes the output response of the power time-slice sequence at the current time step after the combined effects of the basic linear term, spline term, and time modulation. Basic linear terms; For spline terms; The learnable parameter matrix of the basic linear terms; The learnable parameter matrix for the spline terms. These are basic activation functions such as SiLU; This represents the spline basis function expansion for each input dimension.

8. The power energy supply and dispatch forecasting method as described in claim 7, characterized in that, In step S4, the time-modulated neural network includes an adaptive grid update mechanism, which generates grid nodes. The expression for this mechanism is: in, This is an adaptive grid node sequence used to adaptively determine the grid node positions based on the actual distribution of the input features. For the number of grid segments, For quantile sampling index, For the first After sorting the dimensional features, in the position The quantile values ​​obtained at the location.

9. The power energy supply and dispatch forecasting method as described in claim 8, characterized in that, The expression for generating a uniform grid node sequence using a uniform grid is: in, A uniform grid node sequence, used to provide reference grid nodes evenly distributed within the boundary interval. It is an equally spaced grid index; and These are all boundary values, and their expressions are: in, This is the boundary expansion amount, used to expand the value range of the input feature when constructing a uniform grid node sequence. For the first The maximum value after sorting the features. For the first The minimum value after sorting the features.

10. The power energy supply and dispatch forecasting method as described in claim 9, characterized in that, In step S4, the grid node sequence and the uniform grid node sequence are linearly weighted and fused to obtain the fused grid node sequence, the expression of which is: in, To merge the grid node sequences, This is the mesh smoothing fusion factor.

Images