A day-ahead scheduling method, device, system and storage medium of a power system

Abstract

The application discloses a day-ahead scheduling method, device and system of a power system and a storage medium, and belongs to the technical field of new energy system control. The day-ahead scheduling method of the power system firstly obtains future online unit capacity by using a security constrained unit commitment model. Such a design can accurately describe unit start-stop, ramping, backup and line safety constraints. Then, multiple historical electricity price sub-sequences and a key feature matrix are input into a two-stage electricity price prediction model to obtain final electricity price prediction results. Such a design can ensure the accuracy of the electricity price prediction results while meeting the timeliness requirements of power grid operation. Finally, the final electricity price prediction results and the future online unit capacity are used for day-ahead scheduling of the power system. That is, the electricity price prediction results and the future unit capacity are embedded in the day-ahead scheduling of the power system, which can improve the timeliness of scheduling decisions and the safety constraint matching capability in the scenario of high proportion of new energy access, while improving the new energy consumption capability.

Description

Technical Field

[0001] This application belongs to the field of new energy system control technology, and more specifically, relates to a day-ahead dispatching method, device, system and storage medium for a power system. Background Technology

[0002] With a high proportion of new energy sources being integrated into the power system, the randomness and volatility of power sources such as wind and solar power have significantly increased, and net load exhibits rapid changes on a day-ahead to intraday scale. When formulating day-ahead operating plans, the power grid dispatch center needs to simultaneously consider system power balance, reserve capacity, unit ramp-up capabilities, minimum start-up and shutdown times, and transmission line safety constraints. The calculation efficiency and accuracy of the dispatch plan directly affect the actual operating status of the power grid. If the dispatch plan fails to respond promptly to new energy output and load fluctuations, it will directly lead to an increase in the risk of wind and solar curtailment, restricting the power grid's capacity to absorb new energy.

[0003] For day-ahead dispatching needs in scenarios with a high proportion of renewable energy, existing mainstream technical approaches each have significant limitations: Safety-constrained unit combination models can accurately describe unit start-up, shutdown, ramp-up, reserve, and line safety constraints, meeting the requirements for dispatching schemes to adapt to physical operating rules. However, in large-scale system scenarios, the computational burden of repeatedly solving complete safety-constrained unit combination models is heavy, failing to support rolling verification and rapid auxiliary decision-making for day-ahead dispatching, and making it difficult to meet the timeliness requirements of grid operation. While data-driven models can be used to construct day-ahead dispatching schemes with faster computation speeds and meet basic timeliness requirements, these models typically ignore core physical factors such as online unit capacity, line power flow constraints, and reserve requirements. The output results cannot be directly adapted to dispatching execution stages such as unit control, AGC regulation, or energy storage charging and discharging, ultimately failing to respond promptly to renewable energy output fluctuations and failing to fundamentally solve the problem of insufficient grid absorption capacity for high-proportion renewable energy.

[0004] In summary, existing day-ahead dispatch schemes for power systems generally fail to balance the operational efficiency of grids with a high proportion of renewable energy and the capacity for renewable energy absorption, necessitating improvements to the technical approaches of existing dispatch schemes. Summary of the Invention

[0005] In view of the above-mentioned defects or improvement needs of the prior art, this application provides a day-ahead dispatching method, apparatus, system and storage medium for power systems, which solves the technical problem that the existing day-ahead dispatching schemes for power systems are unable to simultaneously take into account the operational efficiency of high-proportion renewable energy grids and the renewable energy absorption capacity.

[0006] To achieve the above objectives, according to one aspect of this application, a day-ahead dispatching method for a power system is provided, comprising: S1: Input the current multi-source operation data of the power system into the safety-constrained unit combination model to obtain the future online unit capacity corresponding to each time period in the future scheduling cycle; S2: Input the target historical electricity price subsequences and target key feature matrices at multiple frequencies corresponding to the multi-source operational data into the two-stage electricity price prediction model to obtain the final electricity price prediction result; the two-stage electricity price prediction model includes: a historical time series feature extraction network and a future external feature fusion network; the historical time series feature extraction network is used to input the target historical electricity price subsequences and output the preliminary electricity price prediction result; the future external feature fusion network is used to input the preliminary electricity price prediction result and the target key feature matrix and output the final electricity price prediction result; The features in the target key feature matrix include the target historical electricity price subsequence, and also include at least one of the following: future online unit capacity, system load forecast data, new energy output forecast data, hydropower planned output, and tie line planned power. S3: Use the final electricity price forecast and future online unit capacity to perform day-ahead dispatch of the power system.

[0007] Furthermore, the historical time series feature extraction network is used to sequentially transform the input target historical electricity price subsequence into: a one-dimensional time series to a two-dimensional time image, a mixed fluctuation feature to a trend feature and a periodic feature, a single-scale feature to a multi-scale fusion feature, and a multi-period representation to an adaptive weighted time series representation, thereby obtaining the preliminary electricity price prediction result.

[0008] Furthermore, the historical time-series feature extraction network comprises the following sequentially connected components: A multi-resolution temporal imaging layer is used to convert the input historical electricity price subsequence into a multi-resolution two-dimensional temporal image; The temporal image decomposition layer is used to separate the periodic fluctuation features and long-term trend features of multi-resolution two-dimensional temporal images to obtain seasonal image features and trend image features. A multi-scale blending layer is used to perform cross-scale feature aggregation on the seasonal image features and the trend image features to obtain multi-scale blended features; A multi-resolution fusion layer is used to adaptively fuse the multi-scale fusion features to obtain the preliminary electricity price prediction result.

[0009] Furthermore, the future external feature fusion network includes: The first basic learner layer is used to input the target key feature matrix and output a first preliminary prediction result. The first preliminary prediction result is used to characterize the global trend and long-term dependency features in the input target key feature matrix. The second basic learner layer is used to input the target key feature matrix and output a second preliminary prediction result. The second preliminary prediction result is used to characterize the local fluctuation pattern and short-term spatial dependence features in the input target key feature matrix. The meta-learner layer is used to receive the first preliminary prediction result and the second preliminary prediction result, and adaptively learn the weight allocation relationship between the first base learner layer and the second base learner layer under different power grid operation scenarios, thereby generating the final electricity price prediction result.

[0010] Furthermore, both the first basic learner layer and the second basic learner layer sequentially include a first feedforward neural network, a convolutional neural network, and a second feedforward neural network.

[0011] Furthermore, before S1, the method includes: S01: acquiring multi-source operating data of the power system; the multi-source operating data includes historical real-time electricity price sequences and planned load output data; the planned load output data includes: system load forecast data, new energy output forecast data, hydropower planned output, and tie-line planned power. Before S2, the process also includes: S02: decomposing the historical real-time electricity price sequence into multiple initial historical electricity price sub-sequences at different frequencies, and selecting a portion of them as target historical electricity price sub-sequences; constructing the target key feature matrix using the target historical electricity price sub-sequences, the future online unit capacity, and the planned load output data.

[0012] Further, S02 includes: The future online unit capacity, the system load forecast data, the new energy output forecast data, the planned hydropower output, and the planned power of the tie line are used as candidate feature data. The maximum mutual information coefficient is used to measure the dependency relationship between each of the initial historical electricity price subsequences, each of the candidate feature data and the target electricity price; the target historical electricity price subsequence and the target feature data are selected according to the dependency relationship, thereby constructing the target key feature matrix.

[0013] According to another aspect of this application, a day-ahead dispatching device for a power system is provided, comprising: The capacity determination module is used to input the current multi-source operating data of the power system into the safety-constrained unit combination model to obtain the future online unit capacity corresponding to each time period in the future scheduling cycle. The electricity price prediction module is used to input the target historical electricity price subsequences at multiple frequencies and the target key feature matrix into a two-stage electricity price prediction model to obtain the final electricity price prediction result. The two-stage electricity price prediction model includes: a historical time series feature extraction network and a future external feature fusion network. The historical time series feature extraction network is used to input the target historical electricity price subsequences and output a preliminary electricity price prediction result. The future external feature fusion network is used to input the preliminary electricity price prediction result and the target key feature matrix and output the final electricity price prediction result. The features in the target key feature matrix include the target historical electricity price subsequence, and also include at least one of the following: future online unit capacity, system load forecast data, new energy output forecast data, hydropower planned output, and tie line planned power. The day-ahead dispatch module is used to perform day-ahead dispatch of the power system using the final electricity price forecast and the future online unit capacity.

[0014] According to another aspect of this application, a day-ahead dispatching system for a power system is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the day-ahead dispatching method for the power system.

[0015] According to another aspect of this application, a computer-readable storage medium is provided, on which a computer program is stored, characterized in that, when the computer program is executed by a processor, it implements the steps of the day-ahead dispatching method of the power system.

[0016] Overall, compared with the prior art, the above-described technical solutions conceived in this application can achieve the following beneficial effects: (1) This application provides a day-ahead dispatching method for a power system. First, the future online unit capacity is obtained using a safety-constrained unit combination model. The future online unit capacity determined in this way can accurately describe unit start-up, shutdown, ramp-up, reserve, and line safety constraints. Then, multiple target historical electricity price subsequences and target key feature matrices are input into a two-stage electricity price prediction model to obtain the final electricity price prediction result. The two-stage electricity price prediction model can ensure the accuracy of the electricity price prediction result while meeting the timeliness requirements of power grid operation. Finally, the day-ahead dispatching of the power system is carried out using the final electricity price prediction result and the future online unit capacity. That is, the electricity price prediction result and the future unit capacity are embedded in the day-ahead dispatching of the power system, which can improve the timeliness and safety constraint matching capability of dispatching decisions in scenarios with a high proportion of new energy access, while reducing unnecessary unit adjustments, thereby improving the new energy absorption capacity.

[0017] (2) The historical time series feature extraction network described in this scheme utilizes a multi-resolution time imaging layer, a time image decomposition layer, a multi-scale mixing layer, and a multi-resolution mixing layer to sequentially transform the input target historical electricity price subsequence into: a one-dimensional time series to a two-dimensional time image, a mixed fluctuation feature to a trend feature and a periodic feature, a single-scale feature to a multi-scale fusion feature, and a multi-period representation to an adaptive weighted time series representation, thereby obtaining the preliminary electricity price prediction result. This design, considering that the electricity price sequence simultaneously possesses trend, periodicity, abrupt change, and multi-time scale coupling characteristics, has the advantage of fully exploring the evolution law of historical electricity prices at different frequencies and periods, and improving the ability of the preliminary electricity price prediction result to characterize peak electricity prices, off-peak electricity prices, and periodic fluctuations.

[0018] (3) The future external feature fusion network described in this scheme includes: a first basic learner layer, a second basic learner layer, and a meta-learner layer. This design takes into account the nonlinear, scenario-specific, and time-varying coupling relationship of the impact of external operating factors such as system load, new energy output, tie-line power, and unit online capacity on electricity prices. The advantage is that different types of external features can be extracted by multiple basic learners, and the prediction contributions of each basic learner can be adaptively fused by the meta-learner, thereby improving the generalization and stability of electricity price prediction results under different operating scenarios.

[0019] (4) In this scheme, the basic learner layer includes a first feedforward neural network, a convolutional neural network, and a second feedforward neural network in sequence. This design takes into account that the target key feature matrix contains both nonlinear mapping relationships across features and local correlation relationships between adjacent time periods and adjacent features. The advantage is that it can first complete the high-dimensional feature mapping through the feedforward neural network, then capture the local fluctuation pattern through the convolutional neural network, and finally output a compact prediction representation through the feedforward neural network, thereby enhancing the basic learner's ability to express complex external features.

[0020] (5) In this scheme, the historical real-time electricity price sequence is decomposed into multiple initial historical electricity price subsequences at different frequencies, and some of them are selected as target historical electricity price subsequences; the target key feature matrix is ​​constructed using the target historical electricity price subsequences, the future online unit capacity, and the planned load output data. This design takes into account that the formation of electricity prices is simultaneously affected by factors such as historical price inertia, system supply and demand status, and unit available capacity. The advantage is that it can uniformly input the multi-frequency fluctuation information of electricity prices and the future power grid operation boundary information into the prediction model, making the prediction results more consistent with the physical operation constraints and market clearing rules under the day-ahead dispatch scenario.

[0021] (6) In this scheme, the maximum mutual information coefficient is used to measure the dependency relationship between each initial historical electricity price subsequence, each candidate feature data and the target electricity price; the target historical electricity price subsequence and the target feature data are selected according to the dependency relationship, thereby constructing the target key feature matrix. This design, considering the existence of redundant variables, weakly correlated variables and key variables that are nonlinearly correlated with the target electricity price in the candidate features, has the advantage of being able to filter out invalid features that contribute little to the electricity price prediction and retain key features with strong nonlinear correlation, thereby reducing the model input dimension, reducing noise interference and improving performance. Attached Figure Description

[0022] Figure 1 This is a flowchart of a day-ahead dispatching method for a power system provided in an embodiment of this application.

[0023] Figure 2 This is a schematic diagram of the structure of a two-stage electricity price prediction model provided in an embodiment of this application.

[0024] Figure 3 This is a schematic diagram of a day-ahead dispatching process for a power system using the final electricity price forecast and future online unit capacity, provided in one embodiment of this application. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application. Furthermore, the technical features involved in the various embodiments described below can be combined with each other as long as they do not conflict with each other.

[0026] Example 1 This embodiment provides a day-ahead dispatching method for a power system, such as... Figure 1 As shown, it includes: S1-S3.

[0027] S1: Input the current multi-source operating data of the power system into the safety-constrained unit combination model to obtain the future online unit capacity corresponding to each time period within the future dispatch cycle. The multi-source operating data includes historical real-time electricity price sequences, system load forecast data, renewable energy output forecast data, planned hydropower output, planned tie-line power, thermal power unit operating status, and unit output upper and lower limits. This multi-source operating data comes from at least the SCADA measurement system, energy management system, renewable energy power plant power forecast system, and market clearing data platform.

[0028] Specifically, the safety-constrained unit combination model is used to simulate the unit combination and market clearing process of the power system in the target area during the day-ahead dispatch phase. The safety-constrained unit combination model aims to minimize the total system operating cost. Under the condition of satisfying the physical operating constraints of the power system, it solves for the unit start-up and shutdown status, active power output, and available online generating capacity of the system at each time period, thus forming online capacity characteristics that characterize the physical supply and demand state of the power grid. Its objective function is to minimize the total system operating cost: In the formula, For the unit i During the period t The operating status indicator variable (a value of 1 indicates power-on, and 0 indicates power-off); This corresponds to the actual power generation output; and Characterizing the units respectively i The fixed operating cost coefficient and marginal generation cost; The set of generator sets participating in the dispatch; To optimize the time period index set within the period.

[0029] The safety-constrained unit combination model must at least satisfy system power balance constraints, reserve capacity constraints, unit output upper and lower limit constraints, unit minimum start-up and shutdown time constraints, unit ramp rate constraints, and transmission line safety constraints. The transmission line safety constraints, based on the power transfer distribution factor, describe the impact of changes in injected power from units on line power flow, ensuring that the online unit capacity and adjustable capacity match the physical operating state of the target area's power grid. The mathematical descriptions of each constraint are as follows: ; In the formula, For time period t System load demand (unit: MW); and The units i The lower and upper limits of output; and For the unit i Minimum continuous start-up time and minimum continuous shutdown time (unit: h); For the unit i The upper limit of the single-time power regulation rate (unit: MW / h); For the line l Transmission capacity limit (unit: MW); The power transfer distribution factor reflects the power transfer distribution of the unit. i Injected power changes on the line l Sensitivity to the impact of power flow. The above model is optimized and solved using commercial mathematical programming solvers (such as CPLEX or Gurobi) to obtain the optimal start-up and shutdown scheme. Then, calculate the time period using the following formula. t Available online power generation capacity : ; The aforementioned online unit capacity, as a physical mechanism feature, is input into the second-stage future external feature fusion network along with system load forecasting, new energy output forecasting, planned hydropower output, planned tie-line power, and the preliminary electricity price forecast results from the first stage of the dual-stage integrated deep learning model.

[0030] As an optional implementation, before S1, the method further includes: S01: acquiring multi-source operating data of the power system; the multi-source operating data includes historical real-time electricity price sequences and planned load output data; the planned load output data includes: system load forecast data, new energy output forecast data, planned hydropower output, and planned tie-line power; before S2, the method further includes: S02: decomposing the historical real-time electricity price sequence into multiple initial historical electricity price subsequences at different frequencies, and selecting a portion as the target historical electricity price subsequence. Specifically, a variational mode decomposition algorithm is used to decompose the historical real-time electricity price sequence, and the number of modes parameter of the variational mode decomposition algorithm is optimized using the center frequency observation method. A target key feature matrix is ​​constructed using the target historical electricity price subsequence, future online unit capacity, and planned load output data.

[0031] As an optional implementation, the variational mode decomposition algorithm is used to analyze the historical electricity price time series. Adaptive decoupling results in several quasi-orthogonal modal components with sparse spectral characteristics. In specific implementation, the number of modes to be decomposed is set. K (Determined through the center frequency observation method). VMD achieves decomposition by solving the following constrained variational problem: .

[0032] The constraints are as follows: In the formula, The original electricity price signal to be decomposed; For the first n One intrinsic mode function; For the first n The center frequency corresponding to each mode (unit: rad / s); i The imaginary unit; For Dirac impulse function; symbol This represents the convolution operation; Describes the space of square-integrable functions Norm, defined as To solve the above constrained optimization problem, a quadratic penalty factor is introduced. (Balance parameter, usually taken as 2000) and Lagrange multipliers Construct the augmented Lagrange function: .

[0033] Subsequently, the alternating direction multiplier method was used to analyze the modal components. Center frequency and the vehicle Perform alternating iterative optimization. Optimize when the convergence criterion is met. ( The iteration terminates when the preset tolerance is reached. Finally, the result is... N A bandwidth-constrained intrinsic mode subsequence Each component contains the fluctuation characteristics and energy distribution of the original electricity price signal at different time scales.

[0034] As an optional implementation, S02 includes: using future online unit capacity, system load forecast data, new energy output forecast data, planned hydropower output, and planned tie-line power as candidate feature data; using the maximum mutual information coefficient to measure the dependency relationship between each initial historical electricity price subsequence, each candidate feature data, and the target electricity price; selecting the target historical electricity price subsequence and target feature data according to the dependency relationship, thereby constructing a target key feature matrix.

[0035] Specifically, for the initial feature set composed of historical mode components obtained through VMD decomposition and other multidimensional features (such as load and renewable energy output), the maximum mutual information coefficient is applied for feature selection; the maximum mutual information coefficient between each feature and the target electricity price is calculated as follows: ; In the formula, As characteristic variables, Target electricity price; and Each is a variable and The number of columns and rows in the grid; A function of sample size (usually taken as...) , (sample size) For variables and Mutual information between them, through statistical analysis. The sample frequency within each cell of the grid is calculated.

[0036] Then, based on the calculated MIC values, all features are sorted in descending order, and a feature selection threshold is set. Or select directly before K One feature, where the MIC value is below a threshold. Or not entered in the sorting KThe features that are strongly correlated with electricity prices are removed; finally, the key feature subset that is strongly correlated with electricity prices (including linear and nonlinear relationships) is retained to form the target key feature matrix.

[0037] S2: Input the target historical electricity price sub-sequences and target key feature matrices at multiple frequencies corresponding to multi-source operational data into the two-stage electricity price prediction model to obtain the final electricity price prediction result. The two-stage electricity price prediction model includes: a historical time-series feature extraction network and a future external feature fusion network. The historical time-series feature extraction network is used as input for the target historical electricity price sub-sequences and outputs preliminary electricity price prediction results. The two-stage electricity price prediction model, such as... Figure 2 As shown, the method does not perform mathematical fitting on the electricity price sequence in isolation. Instead, it generates a future electricity price prediction vector for a preset period, such as 24 hours, driven by the physical state characteristics generated by the safe-constrained unit combination, load and renewable energy forecast characteristics, tie-line planned power, and VMD multi-scale electricity price modal components. This vector is then used to assist in the generation and verification of day-ahead dispatch schemes. A future external feature fusion network is used as input for the preliminary electricity price prediction results and the target key feature matrix, and outputs the final electricity price prediction results. The features in the target key feature matrix include the target historical electricity price subsequence, as well as at least one of the following: future online unit capacity, system load forecast data, renewable energy output forecast data, planned hydropower output, and planned tie-line power.

[0038] As an optional implementation, the historical time series feature extraction network is used to sequentially transform the input target historical electricity price subsequence into: a one-dimensional time series to a two-dimensional time image, a mixed fluctuation feature to a trend feature and a periodic feature, a single-scale feature to a multi-scale fusion feature, and a multi-period representation to an adaptive weighted time series representation, thereby obtaining preliminary electricity price prediction results.

[0039] As an optional implementation, the historical time-series feature extraction network comprises, in sequence, a multi-resolution temporal imaging layer, a temporal image decomposition layer, a multi-scale mixing layer, and a multi-resolution mixing layer. Specifically, the multi-resolution temporal imaging layer converts the input historical electricity price subsequence into a multi-resolution two-dimensional temporal image; the temporal image decomposition layer separates the periodic fluctuation features and long-term trend features of the multi-resolution two-dimensional temporal image to obtain seasonal image features and trend image features; the multi-scale mixing layer aggregates the seasonal image features and trend image features across scales to obtain multi-scale mixed features; and the multi-resolution mixing layer adaptively fuses the multi-scale mixed features to obtain a preliminary electricity price prediction result.

[0040] Specifically, the historical time-series feature extraction network employs the TimeMixer model to extract time-series features from historical electricity price series in the first stage. The TimeMixer model comprises a multi-resolution time imaging layer, a time image decomposition layer, a multi-scale mixing layer, and a multi-resolution mixing layer connected in sequence. The output of each layer serves as the input to the next, enabling the historical electricity price series to undergo a progressive transformation from one-dimensional time series to two-dimensional time image, from mixed fluctuation features to trend and periodic features, from single-scale features to multi-scale fusion features, and from multi-period representation to adaptive weighted time-series representation.

[0041] The multi-resolution temporal imaging layer is used to convert the input historical electricity price subsequences into multi-resolution two-dimensional temporal images. Specifically, this layer takes the multiple historical electricity price subsequences obtained from the aforementioned decomposition process as input and performs a Fast Fourier Transform to extract the dominant periodic component. Subsequently, based on the dominant periodic component, each historical electricity price subsequence is reconstructed into a two-dimensional temporal image according to different time scales. The output of this layer is a multi-resolution temporal image with multiple different dominant periods. The multi-resolution two-dimensional temporal image serves as the input for subsequent temporal image decomposition layers.

[0042] The temporal image decomposition layer is used to separate periodic fluctuation features and long-term trend features from multi-resolution temporal images. Specifically, this layer takes the multi-resolution two-dimensional temporal image output from the aforementioned multi-resolution temporal imaging layer as input, and performs attention modeling along both the temporal and periodic directions of the two-dimensional temporal image to obtain dual-axis attention weights reflecting continuous temporal evolution and periodic repetition relationships. Subsequently, based on the dual-axis attention weights, the periodic fluctuation components and trend evolution components in the multi-resolution two-dimensional temporal image are separated. The output of this layer includes seasonal image features and trend image features. The seasonal image features and trend image features are used together as input to the subsequent multi-scale mixing layer.

[0043] A multi-scale fusion layer is used to aggregate seasonal and trend image features across scales. Specifically, this layer takes the seasonal and trend image features output from the aforementioned temporal image decomposition layer as input and extracts features from them using hierarchical convolution operators with different receptive fields. Subsequently, this layer fuses the extracted seasonal and trend features to form a multi-scale fusion feature. The output of this layer is the multi-scale fusion feature representation. This multi-scale fusion feature representation serves as the input to subsequent multi-resolution fusion layers.

[0044] A multi-resolution fusion layer is used to adaptively fuse the multi-scale fusion feature representation output by the multi-scale fusion layer. Specifically, this layer takes the multi-scale fusion feature representation output by the aforementioned multi-scale fusion layer as input and combines it with the dominant periodic components obtained by Fast Fourier Transform in the multi-resolution temporal imaging layer to determine the weights of different dominant periodic components in the fusion process. Subsequently, this layer performs weighted fusion of the multi-scale fusion feature representation according to the weights to obtain the final historical electricity price time-series characteristic representation. The final historical electricity price time-series characteristic representation is then used to generate the first-stage preliminary electricity price prediction results via a feedforward neural network.

[0045] As an optional implementation, the future external feature fusion network includes: a first basic learner layer, used to input a target key feature matrix composed of future load forecasts, new energy output forecasts, planned hydropower, planned tie-line power, and online unit capacity, and output a first preliminary prediction result, extracting global trends and long-term dependency features from the input target key feature matrix. Alternatively, both the first and second basic learner layers may sequentially include a first feedforward neural network, a convolutional neural network, and a second feedforward neural network. A three-layer feedforward neural network is used to extract global trends and long-term dependency features from the input key feature matrix through a nonlinear activation function. The second basic learner layer is used to input a target key feature matrix composed of future load forecasts, new energy output forecasts, planned hydropower, planned tie-line power, and online unit capacity, and output a second preliminary prediction result, capturing local fluctuation patterns and short-term spatial dependencies in the input target key feature matrix. A hybrid convolutional neural network-feedforward neural network is used, where the convolutional layer captures local fluctuation patterns and short-term spatial dependencies in the input features, and the subsequent feedforward neural network is responsible for integrating local features. The meta-learner layer receives the outputs of the first and second basic learner layers and adaptively learns the weight distribution relationships of the first and second basic learner layers under different power grid operation scenarios to generate the final electricity price prediction result. Specifically, the meta-learner layer receives the outputs of each basic learner, uses a three-layer feedforward neural network as the meta-learner, and adaptively learns the weight distribution relationships of different basic learners under different power grid operation scenarios to generate a real-time electricity price prediction vector for the next 24 hours. This real-time electricity price prediction vector for the next 24 hours is further input into the day-ahead dispatch system to assist in generating generator start-up and shutdown plans, active power output adjustment suggestions, and energy storage charging and discharging power suggestions.

[0046] Furthermore, the future external feature fusion network adopts a three-stage strategy: In the first stage, the parameters of the basic learners are frozen, and only the meta-learner layer is trained, so that the meta-learners can initially learn the combination relationship between the outputs of different basic learners; in the second stage, the connection layer parameters of the basic learners are unfrozen and fine-tuned, so that the basic learners can adapt to the input feature space composed of safety-constrained unit combinations, power grid physical operation characteristics, and VMD modal components; in the third stage, end-to-end global parameter optimization is performed to improve the model's prediction stability and scheduling applicability under typical operating scenarios such as high-proportion renewable energy fluctuations, supply and demand tensions, and limited transmission capacity.

[0047] S3: Day-ahead dispatch of the power system using the final electricity price forecast and future online unit capacity.

[0048] Specifically, the future time period, such as the 24-hour electricity price forecast vector, is jointly verified with constraints on online unit capacity, adjustable capacity, reserve demand, and line transmission capacity. When the electricity price forecast vector shows a peak and the adjustable capacity of online units is insufficient, a scheduling verification result is generated to increase reserve capacity or adjust the output of adjustable units. When the electricity price forecast vector indicates the risk of limited renewable energy consumption, a verification result to reduce the minimum technical output of thermal power units or an energy storage charging power command is generated. When the electricity price forecast vector indicates the risk of supply and demand tension, an energy storage discharging power command or an adjustable unit active power output enhancement curve is generated. Furthermore, the unit start-up and shutdown verification results and the unit active power output adjustment curve are sent to the generator controller or automatic generation control system, and the energy storage charging and discharging power command is sent to the energy storage converter. Thus, the electricity price forecast vector, as an intermediate technical quantity in the day-ahead scheduling optimization and equipment control of the power system, participates in the formation of specific physical equipment control actions.

[0049] The following section uses a regional power system as an example, collecting hourly real-time electricity price data and related power grid operation characteristic data from January 1, 2025 to December 31, 2025. The data includes system load, renewable energy output, planned hydropower output, planned tie-line power, and online unit capacity calculated from a safety-constrained unit combination model. After data preprocessing, the data is divided into training, validation, and test sets in a 0.7:0.15:0.15 ratio according to time sequence.

[0050] The model hyperparameters were determined through grid search, and the number of VMD modes K was 7. After training, the electricity price prediction vector for the next 24 hours was evaluated on the test set. Evaluation metrics included mean absolute error, root mean square error, weighted average percentage error, and coefficient of determination. It should be noted that in this invention, the prediction error metric is only used as a reliability evaluation indicator for intermediate quantities in dispatch auxiliary decision-making, and is not the only or final technical effect of this invention; this invention is ultimately used to support day-ahead unit start-up and shutdown verification, output adjustment, and energy storage charging and discharging control.

[0051] Model evaluation uses mean absolute error (MAE) (yuan / MWh), root mean square error (RMSE) (yuan / MWh), weighted average percentage error (WMAPE) (%), and symmetric average percentage error (R²). 2 As an indicator.

[0052] This embodiment comprehensively evaluates the model performance on the test set. Table 1 shows a comparison of this method with other deep learning methods. The comparison models include single-stage models such as Multilayer Perceptron (MLP), Convolutional Neural Network (CNN), Bidirectional Long Short-Term Memory Network (BiLSTM), Temporal Convolutional Network (TCN), Transformer, TimeMixer, and CNN-LSTM, as well as two-stage models such as Temporal Convolutional Network-Ensemble Learning (TCN-EL), Transformer-Ensemble Learning (Transformer-EL), TimeMixer-TCN, TimeMixer-Transformer, and TimeMixer-StaticEL.

[0053] Table 1

[0054] The results of the implementation examples show that the two-stage architecture has better prediction reliability than the single-stage model. While TimeMixer has limited performance as a single-stage model, it complements future external operational features as a historical time-series feature extractor within the two-stage framework of this invention. Dynamic fusion based on meta-learners further reduces errors compared to static weighted fusion. These results demonstrate that the electricity price prediction vector can serve as a reliable intermediate input for day-ahead scheduling auxiliary decision-making.

[0055] In scheduling applications, such as Figure 3 As shown, the predicted electricity price is not output to market participants as a separate trading strategy. Instead, it is input into the day-ahead dispatch verification module along with the online unit capacity and adjustable capacity obtained from the safety-constrained unit combination model. When the predicted electricity price spikes during the evening peak period and the adjustable capacity is insufficient, the day-ahead dispatch verification module generates suggestions to increase reserve capacity or adjust the output of some adjustable units in advance. When the output of renewable energy is high at midday and the predicted electricity price drops significantly, the day-ahead dispatch verification module generates suggestions to charge energy storage or reduce the minimum technical output of thermal power units. These suggestions can be sent to the control systems of the corresponding equipment after confirmation by the dispatcher.

[0056] Example 2 This embodiment provides a day-ahead dispatching device for a power system, including: a capacity determination module, an electricity price prediction module, and a day-ahead dispatching module.

[0057] The capacity determination module is used to input the current multi-source operating data of the power system into the safety-constrained unit combination model to obtain the future online unit capacity corresponding to each time period within the future scheduling cycle.

[0058] The electricity price prediction module is used to input the target historical electricity price subsequences and the target key feature matrix at multiple frequencies into a two-stage electricity price prediction model to obtain the final electricity price prediction result. The two-stage electricity price prediction model includes: a historical time series feature extraction network and a future external feature fusion network. The historical time series feature extraction network is used to input the target historical electricity price subsequences and output the preliminary electricity price prediction result. The future external feature fusion network is used to input the preliminary electricity price prediction result and the target key feature matrix and output the final electricity price prediction result. Among them, the features in the target key feature matrix include the target historical electricity price subsequences, and also include at least one of the following: future online unit capacity, system load prediction data, renewable energy output prediction data, hydropower planned output, and tie-line planned power.

[0059] The day-ahead dispatch module is used to perform day-ahead dispatch of the power system using the final electricity price forecast and the future online unit capacity.

[0060] Example 3 This embodiment provides a day-ahead dispatching system for a power system, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the day-ahead dispatching method for the power system described above.

[0061] The day-ahead dispatching system of a power system can be a computing device such as a desktop computer, laptop, handheld computer, or cloud server. The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Memory can be used to store computer programs and / or modules. The processor implements various functions of the electronic device by running or executing the computer programs and / or modules stored in memory, and by accessing data stored in memory.

[0062] Example 4 This embodiment provides a computer-readable storage medium storing a computer program thereon, characterized in that the computer program, when executed by a processor, implements the steps of the above-described day-ahead dispatching method for the power system.

[0063] Specifically, the memory may include high-speed random access memory, as well as non-volatile memory, such as hard disks, RAM, plug-in hard disks, smart media cards (SMC), secure digital (SD) cards, flash cards, at least one disk storage device, flash memory device, or other volatile solid-state storage devices.

[0064] Example 5 This application provides a computer program product or computer program that includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the steps of the method described in the above embodiments of this application.

[0065] The technical features of the embodiments described above can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification. It should be noted that the terms "in one embodiment," "for example," and "as in another example" in this application are intended to illustrate the application and are not intended to limit the application.

[0066] The embodiments described above are merely examples of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these modifications and improvements all fall within the protection scope of this application.

Claims

1. A day-ahead dispatching method for a power system, characterized in that, include: S1: Input the current multi-source operation data of the power system into the safety-constrained unit combination model to obtain the future online unit capacity corresponding to each time period in the future scheduling cycle; S2: Input the target historical electricity price subsequences and target key feature matrices at multiple frequencies corresponding to the multi-source operation data into the two-stage electricity price prediction model to obtain the final electricity price prediction result; The two-stage electricity price prediction model includes: a historical time-series feature extraction network and a future external feature fusion network; the historical time-series feature extraction network is used to input the target historical electricity price subsequence and output a preliminary electricity price prediction result; the future external feature fusion network is used to input the preliminary electricity price prediction result and the target key feature matrix and output the final electricity price prediction result. The features in the target key feature matrix include the target historical electricity price subsequence, and also include at least one of the following: future online unit capacity, system load forecast data, new energy output forecast data, hydropower planned output, and tie line planned power. S3: Use the final electricity price forecast results and future online unit capacity to perform day-ahead scheduling of the power system; The historical time series feature extraction network is used to sequentially transform the input target historical electricity price subsequence into: a one-dimensional time series to a two-dimensional time image, a mixed fluctuation feature to a trend feature and a periodic feature, a single scale feature to a multi-scale fusion feature, and a multi-period representation to an adaptive weighted time series representation, to obtain the preliminary electricity price prediction result. The future external feature fusion network includes: The first basic learner layer is used to input the target key feature matrix and output a first preliminary prediction result. The first preliminary prediction result is used to characterize the global trend and long-term dependency features in the input target key feature matrix. The second basic learner layer is used to input the target key feature matrix and output a second preliminary prediction result. The second preliminary prediction result is used to characterize the local fluctuation pattern and short-term spatial dependence features in the input target key feature matrix. The meta-learner layer is used to receive the first preliminary prediction result and the second preliminary prediction result, and adaptively learn the weight allocation relationship between the first base learner layer and the second base learner layer under different power grid operation scenarios, thereby generating the final electricity price prediction result.

2. The day-ahead dispatching method for a power system as described in claim 1, characterized in that, The historical time-series feature extraction network comprises the following sequentially connected components: A multi-resolution temporal imaging layer is used to convert the input historical electricity price subsequence into a multi-resolution two-dimensional temporal image; The temporal image decomposition layer is used to separate the periodic fluctuation features and long-term trend features of multi-resolution two-dimensional temporal images to obtain seasonal image features and trend image features. A multi-scale blending layer is used to perform cross-scale feature aggregation on the seasonal image features and the trend image features to obtain multi-scale blended features; A multi-resolution fusion layer is used to adaptively fuse the multi-scale fusion features to obtain the preliminary electricity price prediction result.

3. The day-ahead dispatching method for a power system as described in claim 1, characterized in that, Both the first basic learner layer and the second basic learner layer sequentially include a first feedforward neural network, a convolutional neural network, and a second feedforward neural network.

4. The day-ahead dispatching method for a power system as described in claim 1, characterized in that, Before S1, the following is also included: S01: Acquire multi-source operation data of the power system; the multi-source operation data includes historical real-time electricity price sequences and planned load output data; the planned load output data includes: system load forecast data, new energy output forecast data, hydropower planned output and tie-line planned power. Before S2, the process also includes: S02: decomposing the historical real-time electricity price sequence into multiple initial historical electricity price sub-sequences at different frequencies, and selecting a portion of them as target historical electricity price sub-sequences; constructing the target key feature matrix using the target historical electricity price sub-sequences, the future online unit capacity, and the planned load output data.

5. The day-ahead dispatching method for a power system as described in claim 4, characterized in that, The S02 includes: The future online unit capacity, the system load forecast data, the new energy output forecast data, the planned hydropower output, and the planned power of the tie line are used as candidate feature data. The maximum mutual information coefficient is used to measure the dependency relationship between each of the initial historical electricity price subsequences, each of the candidate feature data and the target electricity price; the target historical electricity price subsequence and target feature data are selected according to the dependency relationship, thereby constructing the target key feature matrix.

6. A day-ahead dispatching device for a power system, characterized in that, For performing day-ahead dispatching of the power system according to any one of claims 1-5, comprising: The capacity determination module is used to input the current multi-source operating data of the power system into the safety-constrained unit combination model to obtain the future online unit capacity corresponding to each time period in the future scheduling cycle. The electricity price prediction module is used to input the target historical electricity price subsequences at multiple frequencies and the target key feature matrix into a two-stage electricity price prediction model to obtain the final electricity price prediction result. The two-stage electricity price prediction model includes: a historical time series feature extraction network and a future external feature fusion network. The historical time series feature extraction network is used to input the target historical electricity price subsequences and output a preliminary electricity price prediction result. The future external feature fusion network is used to input the preliminary electricity price prediction result and the target key feature matrix and output the final electricity price prediction result. The features in the target key feature matrix include the target historical electricity price subsequence, and also include at least one of the following: future online unit capacity, system load forecast data, new energy output forecast data, hydropower planned output, and tie line planned power. The day-ahead dispatch module is used to perform day-ahead dispatch of the power system using the final electricity price forecast and the future online unit capacity.

7. A day-ahead dispatching system for a power system, comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the day-ahead dispatching method for the power system according to any one of claims 1 to 5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the day-ahead dispatching method for the power system as described in any one of claims 1 to 5.

Images