Time-of-use electricity price generation method based on new energy time sequence output prediction
By integrating renewable energy output data and load data, utilizing improved data dimensionality reduction and clustering models, and combining LSTM models for load forecasting and electricity price period segmentation, and dynamically adjusting electricity prices, the problem of insufficient prediction accuracy and signal distortion of traditional electricity prices under renewable energy access has been solved, thus achieving efficient grid operation and renewable energy consumption.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHONGQING UNIV
- Filing Date
- 2026-02-02
- Publication Date
- 2026-05-08
AI Technical Summary
Traditional electricity price generation methods have failed to effectively adapt to power systems with a high proportion of renewable energy integration, resulting in insufficient prediction accuracy, distorted electricity price signals, and difficulty in achieving flexible adaptation to peak shaving and valley filling as well as fluctuations in renewable energy output.
By integrating historical load data with renewable energy output data, improving the local linear embedding algorithm and kernel principal component analysis for data dimensionality reduction, combining K-means clustering and long short-term memory network models for load forecasting, constructing a time-of-use pricing model, and generating correction coefficients based on renewable energy output deviations to dynamically adjust the electricity price signal.
It has achieved high-precision load forecasting and dynamic electricity price generation, optimized the allocation of power resources, promoted the consumption of new energy sources, and improved the economy and stability of power grid operation.
Smart Images
Figure CN121998722A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of power forecasting technology, and in particular relates to a method for generating time-of-use electricity prices based on the time-series output forecasting of new energy sources. Background Technology
[0002] As the proportion of new energy sources (such as wind power and photovoltaics) in the power system continues to increase, the intermittent and fluctuating characteristics of their output pose serious challenges to the power grid supply and demand balance and the design of electricity pricing mechanisms. Traditional electricity price generation methods mostly rely on a single data source, such as historical load data, and lack dynamic response to the time-series output characteristics of new energy sources, making it difficult to adapt to the power system operation requirements under a high proportion of new energy integration.
[0003] At the load forecasting level, traditional models often directly model high-dimensional load data without considering the impact of multi-dimensional features such as weather, seasons, and holidays. Furthermore, they have poor adaptability to different load patterns, resulting in insufficient forecast accuracy and failing to provide reliable data support for electricity price generation. Regarding time-of-use pricing, existing methods mostly adopt fixed time-of-use rules, failing to fully consider the dynamic characteristics of the load curve and decision-makers' preferences for the length of each time period. This makes it difficult to achieve the goal of "peak shaving and valley filling" and also unable to flexibly adapt to load curve changes caused by fluctuations in renewable energy output.
[0004] Furthermore, discrepancies between forecasted and actual renewable energy output are common, and traditional electricity pricing mechanisms lack effective deviation correction mechanisms, which can easily lead to distorted electricity price signals: when renewable energy output is in excess, electricity prices are not lowered in time to incentivize consumption; when a supply gap occurs, electricity prices are not raised in time to compensate for peak-shaving costs, thereby affecting the safe and stable operation of the power grid.
[0005] To address the aforementioned problems with existing technologies, there is an urgent need to propose a time-of-use electricity price generation method based on the time-series output prediction of new energy sources. Summary of the Invention
[0006] To address the aforementioned technical problems, this invention provides a method for generating time-of-use electricity prices based on the time-series output prediction of new energy sources.
[0007] This invention proposes a time-of-use electricity price generation method based on renewable energy time-series output prediction, comprising the following steps: Historical load data and renewable energy output data are integrated and preprocessed to obtain the raw dataset; The original dataset is subjected to local dimensionality reduction and global dimensionality reduction in sequence to obtain a dimensionality-reduced dataset; Clustering the reduced-dimensional dataset yields several load categories; LSTM models are trained for each load category to output load forecasts for future periods; A time-of-use pricing time-sharing model is constructed. Based on the load forecast results, the time-of-use pricing time-sharing model is solved under a preset duration constraint to obtain a dynamic time-sharing scheme. Based on real-time load data, the load forecast results, and the deviation in new energy output, a correction coefficient is generated. Based on the dynamic time-sharing scheme, the base electricity price for each time period, and the correction coefficient, the final time-of-use electricity price is determined.
[0008] Optionally, the integration and preprocessing of historical load data and renewable energy output data includes: Historical load data and renewable energy output data, including those with weather, seasonal, and holiday characteristics, were collected. Missing values were filled in, outliers were removed, and the time dimension was standardized to a fixed interval to form the original dataset.
[0009] Optionally, the original dataset is subjected to local dimensionality reduction and global dimensionality reduction sequentially to obtain a dimensionality-reduced dataset, including: The LLE algorithm improves the distance between sample points by replacing the traditional Euclidean distance in the LLE algorithm. The improved LLE algorithm is then used to perform local dimensionality reduction on the original dataset, which is then mapped to a higher-dimensional space using a kernel function. The mapped data is then decentralized, and the covariance matrix is calculated, along with the eigenvalues and eigenvectors. Finally, the eigenvectors are sorted by eigenvalue from largest to smallest, and the dimensionality-reduced dataset is output.
[0010] Optionally, the reduced-dimensionality dataset is clustered to obtain several load categories, including: The optimal number of clusters is determined by the silhouette coefficient method, and K-means clustering is performed using Euclidean distance as a metric to classify data with similar characteristics into the same load category.
[0011] Optionally, LSTM models are trained separately for each load category to output load forecasts for future periods, including: LSTM models are built for each load category, and the network parameters of the LSTM models are optimized using the data of the corresponding category. The clustered data are then input into the corresponding optimized LSTM models, and the load prediction results are output.
[0012] Optionally, a time-of-use pricing time-sharing model is constructed. Based on the load forecast results, the time-of-use pricing time-sharing model is solved under a preset duration constraint to obtain a dynamic time-sharing scheme, including: Based on load forecast values, peak and valley membership degrees are calculated at each time point using a semi-trapezoidal membership function. The total duration, minimum continuous duration, and maximum continuous duration of the time period are set as controllable constraints. A time-of-use pricing time period division model is constructed with the goal of minimizing the sum of membership degree distances within the time period. The time-of-use pricing time period division model is solved to obtain the time set corresponding to each time period, forming a dynamic time period division scheme.
[0013] Optionally, solving the time-of-use pricing time-segmentation model further includes: By introducing zero-one auxiliary variables, nonlinear constraints are transformed into zero-one integer programming, and global variables are used to dynamically generate auxiliary variables and constraints to achieve model solution.
[0014] Optionally, the step of generating correction coefficients based on real-time load data, load forecast results, and new energy output deviations includes: Calculate the deviation rate between the predicted output and the actual output of new energy sources. When the actual output is higher than the prediction, the correction coefficient is adjusted downward according to the positive deviation rate. When the actual output is lower than the prediction, the correction coefficient is adjusted upward according to the negative deviation rate. When the deviation is within the preset range, the correction coefficient is taken as the benchmark value.
[0015] Optionally, determining the final time-of-use electricity price includes: The predicted electricity price based on the base price for each time period is multiplied by the correction factor to obtain the final time-of-use electricity price under the real-time supply and demand status.
[0016] The present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method.
[0017] Compared with the prior art, the present invention has the following advantages and technical effects: This invention integrates historical load data with renewable energy output data, utilizes an improved local linear embedding algorithm combined with kernel principal component analysis for data dimensionality reduction, then employs K-means clustering to classify load categories, and trains a long short-term memory network model for each category to perform high-precision load forecasting. Based on the forecast results, a membership function is used to process the load data, and a time-of-use pricing time-sharing scheme is dynamically generated under constraints such as total time period duration and minimum / maximum continuous duration. Furthermore, a correction coefficient is generated in real time based on the renewable energy output forecast deviation rate to dynamically adjust the electricity price signal. This invention effectively solves the problems of insufficient response to output fluctuations, low forecast accuracy, and rigid time-sharing segmentation in the electricity price mechanism under high-proportion renewable energy access. It ensures that when renewable energy output is excessive, the electricity price is lowered to incentivize absorption, and when there is a supply gap, the electricity price is raised to compensate for peak-shaving costs, thus solving the signal distortion problem caused by forecast deviations in traditional electricity pricing. This invention achieves dynamic electricity price generation that accurately reflects the power grid supply and demand status, optimizes power resource allocation, promotes renewable energy absorption, and improves the economy and stability of power grid operation. Attached Figure Description
[0018] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a schematic diagram of the method flow according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the KPCA-K-means-LSTM prediction model framework according to an embodiment of the present invention; Figure 3 This is a flowchart of the time period segmentation model calculation in an embodiment of the present invention. Detailed Implementation
[0019] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.
[0020] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.
[0021] Example 1 like Figure 1 As shown, this embodiment provides a method for generating time-of-use electricity prices based on the time-series output prediction of new energy sources, including the following steps: Historical load data and renewable energy output data are integrated and preprocessed to obtain the raw dataset; The original dataset is subjected to local dimensionality reduction and global dimensionality reduction in sequence to obtain a dimensionality-reduced dataset; Clustering the reduced-dimensional dataset yields several load categories; LSTM models are trained for each load category to output load forecasts for future periods; A time-of-use pricing time-sharing model is constructed. Based on the load forecast results, the time-of-use pricing time-sharing model is solved under a preset duration constraint to obtain a dynamic time-sharing scheme. Based on real-time load data, the load forecast results, and the deviation in new energy output, a correction coefficient is generated. Based on the dynamic time-sharing scheme, the base electricity price for each time period, and the correction coefficient, the final time-of-use electricity price is determined.
[0022] As a specific implementation method, the process of integrating and preprocessing historical load data and renewable energy output data to obtain the raw dataset includes: Data preprocessing involves collecting historical load data containing features such as weather, seasons, and holidays, cleaning the data (filling in missing values and removing outliers), and standardizing the time dimension (e.g., standardizing it to a fixed time interval) to form the original dataset. As a specific implementation method, the process of performing local dimensionality reduction and global dimensionality reduction sequentially on the original dataset to obtain a dimensionality-reduced dataset includes: This dimensionality reduction approach, based on the KPCA-L algorithm, uses an improved LLE method (calculating sample point distances using the shortest path algorithm instead of traditional Euclidean distance) to perform local dimensionality reduction. The locally dimensionality-reduced dataset is then mapped to a higher-dimensional space using a kernel function, making it linearly separable. The mapped data is then decentered, and the covariance matrix is calculated, along with the eigenvalues and eigenvectors. The eigenvectors are sorted in descending order of eigenvalues, and the first r eigenvectors are retained to form a matrix, which is then output as the dimensionality-reduced dataset. The basic idea of the traditional kernel PCA algorithm, which is feasible, is to map linearly inseparable data in a low-dimensional space to a higher-dimensional space using a kernel function, making it linearly separable in the higher-dimensional space. To improve the speed and complexity of dimensionality reduction, the KPCA-L algorithm first preprocesses the original data before mapping it, then uses an improved LLE method to perform local dimensionality reduction on the original data, and finally uses kernel PCA for global dimensionality reduction.
[0023] Traditional LLE methods require calculating the Euclidean distance between sample points to find the k nearest neighbors. While Euclidean distance is simple to calculate in low-dimensional spaces, the complexity in high-dimensional spaces makes it unreliable as a similarity measure between two sample points. Therefore, this embodiment uses the shortest path algorithm to calculate the distance between two sample points, thereby improving the traditional LLE method.
[0024] The KPCA-L algorithm is described in detail below: (1) Perform dimensionality reduction preprocessing on dataset X; (2) After dimensionality reduction, the dataset Y is processed by kernel function to map to a high-dimensional space, making the data linearly separable. Then, the matrix is decentered, the covariance matrix is calculated, and the eigenvalues and eigenvectors are solved. (3) Arrange the eigenvectors in descending order of eigenvalues, retain the first r eigenvectors to form matrix P, and finally output the dimensionality-reduced dataset V.
[0025] As a specific implementation method, the process of clustering the dimensionality-reduced dataset to obtain several load categories includes: Load category clustering based on K-means employs the silhouette coefficient method, selecting the K value with the largest silhouette coefficient (the closer the silhouette coefficient is to 1, the better the clustering effect). Using Euclidean distance as the metric, the dimensionality-reduced dataset is divided into K categories, so that load data with similar features (such as similar weather and power curves) are grouped into the same category. Feasible load forecasting primarily considers factors such as weather, season, and whether the day is a holiday. This embodiment clusters datasets with similar weather and power curves into different categories, and trains the network on each category to obtain optimal parameters. This embodiment uses the K-means method for unsupervised clustering. K-means clustering randomly selects K points as cluster centers, and assigns the remaining samples to different categories based on their Euclidean distance from these cluster centers. The Euclidean distance is shown below: MERGEFORMAT (1) In the formula: x i y i Let X and Y be the i-th variables of samples X and Y, respectively.
[0026] Methods for determining the K value are divided into the elbow method and the silhouette coefficient method. The elbow method has low automation and its efficiency needs to be improved when dealing with massive amounts of data. The silhouette coefficient method determines the K value based on the clustering cohesion and separation. The silhouette coefficients corresponding to different K values are shown in Table 1. Table 1 The silhouette coefficient is typically between -1 and 1, and a larger silhouette coefficient indicates a better clustering effect.
[0027] As a specific implementation method, the process of training LSTM models for each load category and outputting load forecast results for future periods includes: LSTM model training involves constructing a separate LSTM model for each category obtained from K-means clustering, training it using the corresponding category's training set data, and optimizing network parameters (such as input gate, forget gate, and output gate parameters) to adapt to the temporal characteristics of different category loads. By adjusting the LSTM network structure (such as the number of layers and hidden units), the gradient vanishing problem of traditional RNNs is solved, enhancing the ability to capture long-term temporal information. Load forecasting: The forecast set data is processed using the KPCA-L dimensionality reduction and K-means clustering process described above to determine its category. An LSTM model corresponding to the forecast sample category is selected, and the processed forecast sample data is input to generate the load forecasting results.
[0028] The LSTM model is a specific type of recurrent neural network. LSTM can analyze inputs using time series data, effectively conveying and representing information from long time sequences without ignoring useful information from much earlier periods. Furthermore, LSTM can address the vanishing gradient problem in RNNs.
[0029] The computation method of the GRU neural network is as follows: MERGEFORMAT (2) MERGEFORMAT (3) MERGEFORMAT (4) MERGEFORMAT (5) In the formula: It's an input gate. and It is the Gate of Oblivion. It is an output gate.
[0030] The dimensionality-reduced training set is clustered into different categories, and different LSTM model parameters are trained for each category. The prediction set is then divided into categories and matched using the same method. The matching results are then input into the neural network for prediction. The specific process is as follows: Figure 2 As shown.
[0031] High-dimensional datasets can more realistically reflect the actual situation. However, the high dimensionality of the original dataset increases processing time and reduces the practicality of the model. KPCA can reduce the dimensionality of the dataset while preserving data information. The K-means algorithm can classify the dimensionality-reduced data, and more suitable network parameters can be obtained by training different categories of data separately.
[0032] As a specific implementation method, such as Figure 3 As shown, the process of constructing a time-of-use pricing time-sharing model, and solving the model under preset duration constraints based on the load forecast results to obtain a dynamic time-sharing scheme includes: Time-of-use pricing is a key component of time-of-use electricity pricing. Dividing the day into different time periods helps to effectively manage electricity supply and demand. Reasonable time-of-use pricing can better utilize electricity resources, optimize power supply plans, and improve the efficiency and stability of the power grid.
[0033] Since a large load value can affect the solution speed of the time period segmentation model, we consider using a membership function to process it. This is because the membership function only needs to determine the positional relationship of each point on the load curve relative to the highest and lowest points to describe the data characteristics, without relying on specific numerical values.
[0034] For ease of calculation, only the linear portion of the semi-trapezoidal distribution function is used. The peak membership degree at the highest point of the load curve is defined as 1, and the valley membership degree as 0; similarly, the peak membership degree at the lowest point of the load curve is defined as 0, and the valley membership degree as 1. The linear portion of the slightly smaller semi-trapezoidal distribution function is monotonically decreasing, while the linear portion of the slightly larger semi-trapezoidal distribution function is monotonically increasing. Based on the characteristics of these two types of semi-trapezoidal distribution functions, the valley membership degree for the remaining points on the load curve is calculated using the slightly smaller semi-trapezoidal distribution function, and the peak membership degree is calculated using the slightly larger semi-trapezoidal distribution function. Their formulas are as follows: MERGEFORMAT (6) MERGEFORMAT (7) in, Let t be the valley membership degree of the load value; The peak membership degree of the load value at time t; Let be the load value at time t; a be the minimum value on the load curve; and b be the maximum value on the load curve.
[0035] Feasible, controllable total duration constraints within the time period: Decision-makers typically have their own subjective preferences regarding the duration of each time period. Therefore, the total duration of each time period can be defined as follows.
[0036] MERGEFORMAT (8) MERGEFORMAT (9) in, This is the judgment matrix for all times in time period T; The value for time t within time period T is a 0-1 variable, where 0 indicates that time t does not belong to time period T, and 1 indicates that time t belongs to time period T. The interval between adjacent time points; The final moment; This represents the total duration of time period T.
[0037] Since each moment can only be divided into a certain time period, there are constraints: MERGEFORMAT (10) in, This represents the total number of time periods.
[0038] If decision-makers want to classify a certain moment into a certain time period to facilitate other definitions, they can define the judgment value of that time period as 1. For example, the time period in which the maximum load is located is usually defined as the peak period. This allows for a better definition of the duration, minimum continuous duration, and maximum continuous duration of the peak period.
[0039] As is feasible, the total duration of each time period should be constrained by the total duration of all times: MERGEFORMAT (11) in, Total duration for all moments.
[0040] Feasible, controllable minimum continuous duration constraint within a time period: In each time period, an excessively short minimum continuous duration can prevent electricity users from adjusting their electricity consumption strategies in a timely manner, which is detrimental to peak shaving and valley filling of the electricity load. Therefore, policymakers will impose requirements on the minimum continuous duration, which is defined in the following two ways for each time period: (1) The minimum continuous duration is not less than one moment: No additional constraints need to be added; the solution can be obtained directly based on the controllable constraints of the time-of-use electricity price period.
[0041] (2) The minimum continuous duration is not less than 2 or more time points: To ensure that the minimum continuous duration of time period T is not less than First, we should calculate the time interval T from each moment before it. The number of completely consecutive moments at any given moment Defined as follows: MERGEFORMAT (12) Recalculate the time interval T corresponding to each moment before it. The time intervals are completely consecutive and their preceding time intervals are... The moment before it Number of discontinuous moments Defined as follows: MERGEFORMAT (13) The start and end times are considered continuous to ensure that the minimum continuous duration of time interval T is not less than According to the pattern, , Constraints must be met: MERGEFORMAT(14) Feasible, controllable maximum continuous duration constraint within a time period: Excessively long maximum continuous duration in any given time period can lead to electricity users shifting their load to the maximum extent, and may even cause a "peak-valley reversal" phenomenon in electricity load. Therefore, policymakers will impose requirements on the maximum continuous duration, which is defined in the following two ways for each time period: (1) The minimum continuous duration is no more than one moment: Since the maximum continuous duration is no more than one moment (i.e., each continuous duration is one moment), the number of moments in time period T that are not consecutive with their preceding moment is equal to the total number of moments in time period T. Therefore: MERGEFORMAT (15) (2) The minimum continuous duration is no more than 2 or more time intervals: To ensure that the maximum continuous duration of time period T does not exceed Similar to ensuring that the minimum continuous duration is no more than 2 or more time points, the first step should be to statistically analyze the time interval T from each time point to its preceding time point. The number of completely consecutive moments at any given moment Then, calculate the time interval T corresponding to each time point before it. The time intervals are completely consecutive and their preceding time intervals are... The moment before it Number of discontinuous moments , , Define it as follows: MERGEFORMAT (16) MERGEFORMAT (17) The start and end times are considered continuous to ensure that the maximum continuous duration of time interval T does not exceed [a certain value]. According to the pattern, , If they should be equal, then add the following constraint: MERGEFORMAT (18) Implementable objective function: Before considering the objective function for time period division, we must also consider the distance within and between time periods, which is the same as the intra-cluster distance and inter-cluster distance in clustering algorithms. Euclidean distance is used to calculate the distance between the membership degrees of load values at each time point.
[0042] MERGEFORMAT (19) in, This represents the distance between the membership degrees of the load values at time i and time j.
[0043] Therefore, the distance matrix D of the membership degree of the load value at each time point is: MERGEFORMAT (20) Here, the objective function is to minimize the distance within a time period. This ensures that the load values within each time period are closer to each other, forming more compact time periods, as detailed below: MERGEFORMAT (21) in, for Transpose matrix.
[0044] Feasible, model solution: Since the time-segmentation model contains variable products and absolute values, it is converted into a 0-1 integer programming problem to make the model easier to calculate. Therefore, auxiliary variables and constraints are added to adjust the formula.
[0045] First, process the absolute value, then... For example, let: MERGEFORMAT (22) For 0-1 auxiliary variables and It has the following constraints: MERGEFORMAT (23) MERGEFORMAT (24) Next, we process the product of variables, to For example, let: MERGEFORMAT (25) For 0-1 auxiliary variables The following constraints apply: MERGEFORMAT (26) MERGEFORMAT (27) MERGEFORMAT (28) As the minimum and maximum continuous durations increase, it becomes necessary to handle not only the products of variables, but also the products of variables and auxiliary variables, and even the products of auxiliary variables themselves, though the processing methods remain the same. Simultaneously, the number of auxiliary variables and constraints required to convert the model into 0-1 integer programming will increase, making the solution increasingly difficult.
[0046] A feasible alternative is Python's `globals` function, which dynamically creates, accesses, and deletes variables, returning a dictionary containing key-value pairs of all global variables in the current scope. The keys are the variable names, and the values are their corresponding values. By using `globals`, variables in the global scope can be dynamically viewed and modified. Therefore, `globals` effectively addresses the problem of repetitive code writing in time-segmentation models.
[0047] As a specific implementation method, the process of generating correction coefficients based on real-time load data, the load forecast results, and the deviation in renewable energy output includes: Net load related parameters: Net load value: This directly reflects the amount of power that traditional power sources need to handle; Net load volatility: This represents the rate of change of net load per unit time (such as a rapid increase in net load due to a sudden drop in photovoltaic power in the evening). The greater the volatility, the more difficult it is to peak load. Net load peak-to-valley ratio: (The ratio of the maximum to the minimum net load during the period) reflects the degree of fluctuation in the net load curve and is used to adjust the benchmark for flat-period electricity prices; Net load threshold: Peak period net load threshold ( (At times, high-priced peak-shaving power supplies need to be activated); Off-peak net load threshold ( At that time, there was an overcapacity in new energy output, and incentives were needed to encourage its consumption.
[0048] The quantification formula after integrating net load parameters: The flat-section electricity price is not only affected by the output of new energy, but also related to the overall fluctuation intensity (peak-valley ratio) of the net load: the larger the peak-valley ratio of the net load, the higher the demand for system reserve capacity, and the flat-section benchmark electricity price needs to be increased.
[0049] \ MERGEFORMAT (29) The parameter k is the "net load fluctuation compensation coefficient" (0 < k < 0.2), indicating that for every 1 increase in the peak-valley ratio of the net load (i.e., curve fluctuation), the flat-section electricity price is increased by k times.
[0050] The electricity price during the peak period mainly depends on the net load gap (i.e., the part where the net load exceeds the critical value): the higher the net load, the higher the marginal cost of traditional power sources and peak regulation resources, and the greater the increase in electricity price.
[0051] \MERGEFORMAT (30) The electricity price during the spike period (such as 12:00 - 14:00 in summer) needs to consider both the "high value" and "high volatility" of the net load (such as sudden cloud occlusion causing a sharp drop in photovoltaic power and a rapid spike in the net load), and both jointly push up the peak regulation cost.
[0052] \ MERGEFORMAT (31) is the net load volatility, is the maximum value of the net load on the day, reflects the proportion of the current net load to the peak value.
[0053] The electricity price during the low valley period is affected by the "excess degree" of the net load: when the net load is lower than the critical value (i.e., the output of new energy is excessive and traditional power sources need to reduce output or shut down), the downward range of the electricity price is expanded to encourage consumption.
[0054] \MERGEFORMAT (32) can take 0.8 times the historical average net load of this period (indicating that when it is lower than this value, it enters the "excess state").
[0055] It is feasible to introduce a deviation correction coefficient based on the original time-of-use electricity price formula, and the final electricity price calculation formula is adjusted to: \ MERGEFORMAT (33) Among them, the electricity price based on prediction is the electricity price calculated by the previous formula, and the input parameters are the predicted output and the predicted load ; the deviation correction coefficient is determined by the real-time deviation rate: Positive deviation rate (actual output > predicted output): ( To correct for sensitivity, a value of 0.005~0.01 was used. Negative deviation rate (actual output < predicted output): ; Deviation rate at within, (Ignore minor deviations).
[0056] The dynamic electricity price generation method based on renewable energy time-series output forecasting proposed in this embodiment achieves multiple technical advantages by integrating load forecasting based on KPCA-LK-means-LSTM, a time-of-use pricing period segmentation model considering duration preferences, and a dynamic deviation quantification and price correction mechanism. By leveraging the KPCA-L algorithm to reduce the dimensionality of high-dimensional data, K-means clustering to classify load categories, and combining this with categorical LSTM model training, the accuracy of load forecasting is effectively improved, providing reliable basic data for electricity price generation. By processing load data through membership functions and optimizing the time-of-use pricing time periods based on constraints such as total time period duration and minimum / maximum continuous duration, the electricity pricing mechanism's ability to optimize the allocation of power resources is enhanced, adapting to actual electricity demand. A basic electricity price formula for each time period is constructed using net load parameters, and a real-time correction coefficient is generated by combining the new energy output forecast deviation rate. This dynamically corrects the electricity price signal, ensuring that when there is a surplus of new energy output, the electricity price is lowered to incentivize absorption, and when there is a supply gap, the electricity price is raised to compensate for peak-shaving costs. This solves the signal distortion problem caused by forecast deviations in traditional electricity pricing.
[0057] Example 2 This embodiment also discloses a computer device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method described in Embodiment 1.
[0058] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.
Claims
1. A method for generating time-of-use electricity prices based on the time-series output prediction of new energy sources, characterized in that, Includes the following steps: Historical load data and renewable energy output data are integrated and preprocessed to obtain the raw dataset; The original dataset is subjected to local dimensionality reduction and global dimensionality reduction in sequence to obtain a dimensionality-reduced dataset; Clustering the reduced-dimensional dataset yields several load categories; LSTM models are trained for each load category to output load forecasts for future periods; A time-of-use pricing time-sharing model is constructed. Based on the load forecast results, the time-of-use pricing time-sharing model is solved under a preset duration constraint to obtain a dynamic time-sharing scheme. Based on real-time load data, the load forecast results, and the deviation in new energy output, a correction coefficient is generated. Based on the dynamic time-sharing scheme, the base electricity price for each time period, and the correction coefficient, the final time-of-use electricity price is determined.
2. The method according to claim 1, characterized in that, The integration and preprocessing of historical load data and renewable energy output data includes: Historical load data and renewable energy output data, including those with weather, seasonal, and holiday characteristics, were collected. Missing values were filled in, outliers were removed, and the time dimension was standardized to a fixed interval to form the original dataset.
3. The method according to claim 1, characterized in that, The original dataset is subjected to local and global dimensionality reduction sequentially to obtain a dimensionality-reduced dataset, including: The LLE algorithm improves the distance between sample points by replacing the traditional Euclidean distance in the LLE algorithm. The improved LLE algorithm is then used to perform local dimensionality reduction on the original dataset, which is then mapped to a higher-dimensional space using a kernel function. The mapped data is then decentralized, and the covariance matrix is calculated, along with the eigenvalues and eigenvectors. Finally, the eigenvectors are sorted by eigenvalue from largest to smallest, and the dimensionality-reduced dataset is output.
4. The method according to claim 1, characterized in that, Clustering the dimensionality-reduced dataset yields several load categories, including: The optimal number of clusters is determined by the silhouette coefficient method, and K-means clustering is performed using Euclidean distance as a metric to classify data with similar characteristics into the same load category.
5. The method according to claim 1, characterized in that, LSTM models are trained separately for each load category, and the load forecast results for future periods are output, including: LSTM models are built for each load category, and the network parameters of the LSTM models are optimized using the data of the corresponding category. The clustered data are then input into the corresponding optimized LSTM models, and the load prediction results are output.
6. The method according to claim 1, characterized in that, A time-of-use pricing time-sharing model is constructed. Based on the load forecast results, the time-of-use pricing time-sharing model is solved under preset duration constraints to obtain a dynamic time-sharing scheme, including: Based on load forecast values, peak and valley membership degrees are calculated at each time point using a semi-trapezoidal membership function. The total duration, minimum continuous duration, and maximum continuous duration of the time period are set as controllable constraints. A time-of-use pricing time period division model is constructed with the goal of minimizing the sum of membership degree distances within the time period. The time-of-use pricing time period division model is solved to obtain the time set corresponding to each time period, forming a dynamic time period division scheme.
7. The method according to claim 6, characterized in that, Solving the time-of-use pricing time-of-use model also includes: By introducing zero-one auxiliary variables, nonlinear constraints are transformed into zero-one integer programming, and global variables are used to dynamically generate auxiliary variables and constraints to achieve model solution.
8. The method according to claim 1, characterized in that, The process of generating correction coefficients based on real-time load data, load forecasting results, and new energy output deviations includes: Calculate the deviation rate between the predicted output and the actual output of new energy sources. When the actual output is higher than the prediction, the correction coefficient is adjusted downward according to the positive deviation rate. When the actual output is lower than the prediction, the correction coefficient is adjusted upward according to the negative deviation rate. When the deviation is within the preset range, the correction coefficient is taken as the benchmark value.
9. The method according to claim 1, characterized in that, The determination of the final time-of-use electricity price includes: The predicted electricity price based on the base price for each time period is multiplied by the correction factor to obtain the final time-of-use electricity price under the real-time supply and demand status.
10. A computer device comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method according to any one of claims 1-9.