A semantic source-network-payload-storage collaborative power sequence prediction model training method

Abstract

The application discloses a semantic source-network-load-storage collaborative power sequence prediction model training method, and steps include: step 1) obtaining original data of source network load storage; step 2) dividing variable interval; step 3) generating word coding according to the corresponding relationship between numerical value and variable interval; step 4) integrating the word coding into one-dimensional word sequence; step 5) determining the input data type and output data type of the power sequence prediction model; step 6) constructing multiple sets of training data by using the one-dimensional word sequence; step 7) training the neural network model by using the training data to obtain the power sequence prediction model; and step 8) inputting the word coding sequence into the power sequence prediction model to obtain the power sequence prediction result. Through semantic coding and sequence modeling technology, the application converts power system data into a sequence in the form of natural language, and realizes efficient utilization of data and improvement of model generalization ability.

Description

Technical Field

[0001] This invention relates to the field of power system dispatching, specifically a semantic source-grid-load-storage collaborative power sequence prediction model training method. Background Technology

[0002] In the field of power system dispatching, traditional methods mainly rely on numerical calculations and physical models, such as optimization algorithms based on linear programming, nonlinear programming, or dynamic programming. These methods require accurate mathematical models and large amounts of real-time data. However, when faced with high-dimensional, uncertain source-grid-load-storage (source: generation side; grid: power grid; load: load; storage: energy storage) data, they often suffer from high computational complexity, poor real-time performance, and insufficient generalization ability. With the large-scale integration of renewable energy sources (such as wind power and photovoltaics), the randomness and volatility of the power system have increased significantly, making the limitations of traditional dispatching methods even more prominent.

[0003] The existing technology has the following main problems and disadvantages:

[0004] Low data utilization efficiency: Existing methods usually process raw numerical data directly, lacking understanding and abstraction of data semantics, resulting in low model training efficiency and difficulty in capturing complex relationships between multiple variables.

[0005] Insufficient training data: In real-world power systems, high-quality historical data is limited and difficult to cover all operational scenarios (such as extreme weather and load fluctuations), affecting the model's generalization ability. Traditional data augmentation methods often rely on simple interpolation or noise addition, failing to generate physically meaningful simulated data.

[0006] High computational resource consumption: Traditional numerical optimization methods require iterative calculations, which impose a heavy computational burden when processing large-scale, high-dimensional data, making it difficult to meet real-time scheduling requirements, especially when deployed on edge devices. Difficulty in multi-source data fusion: Complex spatiotemporal coupling relationships exist between multi-dimensional time-series data such as load, wind power, and photovoltaics. Traditional methods (such as independently processing each variable) struggle to effectively capture these relationships, leading to scheduling decision biases. Poor model interpretability: Black-box machine learning methods (such as deep neural networks) lack semantic interpretation, hindering scheduling personnel's understanding and trust.

[0007] Furthermore, existing technologies have attempted to use machine learning methods for scheduling, but these typically rely on large amounts of labeled data, involve complex training processes, and lack effective pre-training mechanisms. For example, some methods directly train recurrent neural networks (RNNs) using numerical sequences, but these models are prone to overfitting and are sensitive to missing data. Therefore, there is an urgent need in this field for an efficient and semantic training method to improve the accuracy, efficiency, and interpretability of power system scheduling. Summary of the Invention

[0008] The purpose of this invention is to provide a semantic source-grid-load-storage collaborative power sequence prediction model training method, comprising the following steps:

[0009] Step 1) Obtain the raw data of source, grid, load, and storage;

[0010] Step 2) Normalize the raw data of source, grid, load and storage, and divide the variable interval according to the normalized value range;

[0011] Step 3) Determine the trends and values ​​of different types of data in the original data, and generate word codes based on the correspondence between the values ​​and variable ranges;

[0012] Step 4) Integrate the word codes into a one-dimensional word sequence according to time order;

[0013] Step 5) Determine the input data type and output data type of the power sequence prediction model; the output data type is either wind power or photovoltaic power.

[0014] Step 6) Based on the input and output data types, construct multiple sets of training data using one-dimensional word sequences;

[0015] Step 7) Train the neural network model using the training data to obtain the power sequence prediction model;

[0016] Step 8) Obtain the current input data and convert it into a word encoding sequence. Input the word encoding sequence into the power sequence prediction model to obtain the power sequence prediction result.

[0017] Furthermore, the raw data for the power generation, grid, load, and energy storage system includes weather, load, wind power, and solar power.

[0018] Furthermore, the raw data of source-grid-load-storage includes actual raw data and simulated raw data;

[0019] The method for generating the simulated original data is as follows:

[0020] Weather data is generated based on historical meteorological data;

[0021] Load power data is generated based on a typical daily load baseline, user behavior correction factors, economic development trend correction factors, and random disturbances, namely:

[0022]

[0023] In the formula, For time period The load power, For hours The corresponding typical daily load benchmark, This refers to the user behavior correction coefficient corresponding to weekdays, weekends, or holidays. This is a seasonal correction factor. This is a correction factor for economic development trends. , and The determination is based on day type, seasonal factors, and economic development factors, respectively. Determined based on load differences between weekdays, weekends, or holidays; Determined based on the impact of monthly, seasonal, or temperature changes on the load; The correction factor is determined based on the annual load growth rate, the level of regional economic development, or historical electricity consumption growth trends. The correction factor can be obtained through historical load data statistics, table lookup, or fitting methods. This is a random disturbance term used to characterize random fluctuations that cannot be directly modeled in load forecasting. The random disturbance term can be a zero-mean random variable, whose variance is determined based on historical load errors or a preset disturbance intensity, and the generated load power is constrained to be no less than zero and not exceeding a preset load upper limit.

[0024] Power output curves are generated using wind power density and solar radiation models;

[0025] The wind power density is shown below:

[0026]

[0027] In the formula, For time period wind power density, air density, For time period wind speed;

[0028] The solar radiation model is shown below:

[0029]

[0030] In the formula, For time period The intensity of solar radiation, The maximum solar radiation intensity, For the current hour, and These are sunrise and sunset times, respectively. This is a weather correction factor; Determined based on factors such as weather type, cloud cover, rainfall, haze, or atmospheric transparency, it is used to characterize the attenuation effect of weather conditions on solar radiation intensity. The range of values ​​is The value is close to 1 in sunny weather, and decreases accordingly in cloudy, rainy, or low-visibility weather. The specific value can be determined by statistical analysis of historical radiation data or by referring to tables based on weather type.

[0031] The energy storage state is generated according to the charging and discharging strategy, that is:

[0032]

[0033] In the formula, For time period The state of charge of energy storage, and Time periods The charging power and discharging power, and These are charging efficiency and discharging efficiency, respectively. For the rated capacity of energy storage, For time step.

[0034] Furthermore, after normalizing the raw data of source-grid-load-storage, the numerical range of the raw data of source-grid-load-storage is 0. ~ 1.

[0035] Furthermore, word encoding includes data type, interval sequence number, and trend; the trend is the change trend of data relative to the previous moment, including rising, falling, and stable.

[0036] Furthermore, in step 4), the words for load, wind power, photovoltaic, and weather at the same time are arranged in a fixed order in the one-dimensional word sequence.

[0037] Furthermore, the one-dimensional word sequence also carries timestamped words.

[0038] Furthermore, the input data for the power sequence prediction model consists of the original data from the power generation, grid, load, and storage systems, excluding the output data.

[0039] Furthermore, in step 6), the steps of constructing multiple sets of training data using one-dimensional word sequences include:

[0040] Step 6.1) Extract consecutive words from a one-dimensional word sequence using the sliding window technique;

[0041] Step 6.2) Randomly hide one word of the same data type as the output word in a series of consecutive words using a masking mechanism, and use the hidden word as the output and the other unhidden words as the input;

[0042] Step 6.3) Repeat steps 6.1)-6.2) until multiple sets of training data are constructed.

[0043] Furthermore, the training objective of the neural network model is to minimize the prediction error, the loss function is the cross-entropy loss function, and the optimization algorithm is Adam or SGD.

[0044] The technical effects of this invention are undeniable, and its beneficial effects are as follows:

[0045] Improve training efficiency: By using semantic encoding and sequence modeling, high-dimensional numerical data is transformed into low-dimensional discrete sequences, reducing model parameters and computational complexity.

[0046] Enhanced generalization ability: By utilizing simulated data and contextual prediction tasks, the model learns general semantic patterns of the power system, achieving a prediction accuracy of 85.5% in scenarios where data is scarce or unseen (such as extreme weather).

[0047] Reduced computational resource consumption: The semantic model only needs to be pre-trained once and can be used multiple times. The inference process is lightweight and suitable for deployment on resource-constrained edge devices, supporting real-time scheduling.

[0048] Improved multi-source data fusion: Interleaved sequences effectively capture the coupling relationship between load, wind power, and photovoltaics, improving the accuracy of scheduling decisions.

[0049] Simplified operation process: The method is highly automated, requiring no manual intervention from data generation to model training, which reduces the operation and maintenance costs and manpower requirements of the scheduling system.

[0050] Enhanced interpretability: Semantic words provide an intuitive representation of data, allowing schedulers to understand the basis of model decisions through word sequences, thus increasing trust.

[0051] Improved robustness: The model has better tolerance to noise and missing data, significantly reduces the error rate (such as word prediction error rate), and maintains stable performance even with poor data quality.

[0052] Supports flexible expansion: The vocabulary and sequence structure can easily incorporate new variables (such as energy storage and electricity prices) to adapt to the evolution of the power system. Attached Figure Description

[0053] Figure 1 A flowchart of a semantic network;

[0054] Figure 2 For training loss curves;

[0055] Figure 3 The semantic vector space is distributed;

[0056] Figure 4 Model training accuracy. Detailed Implementation

[0057] The present invention will be further described below with reference to embodiments, but it should not be construed that the scope of the present invention is limited to the following embodiments. Various substitutions and modifications made based on ordinary technical knowledge and common practices in the art without departing from the above-described technical concept of the present invention should be included within the scope of protection of the present invention.

[0058] Example 1:

[0059] See Figures 1 to 4 A semantic source-grid-load-storage collaborative power sequence prediction model training method includes the following steps:

[0060] Step 1) Obtain the raw data of source, grid, load, and storage;

[0061] Step 2) Normalize the raw data of source, network, load, and storage, and divide the variables into intervals according to the normalized numerical range. The number of intervals for different variables can be set according to the variable fluctuation characteristics, vocabulary size, and validation set prediction error. The number of intervals is determined based on the data distribution, sampling resolution, vocabulary size, and validation set prediction error to balance the preservation of numerical details and the sparsity of training samples.

[0062] Step 3) Determine the trends and values ​​of different types of data in the original data, and generate word codes based on the correspondence between the values ​​and variable ranges;

[0063] Step 4) Integrate the word codes into a one-dimensional word sequence according to time order;

[0064] Step 5) Determine the input data type and output data type of the power sequence prediction model; the output data type is either wind power or photovoltaic power.

[0065] Step 6) Based on the input and output data types, construct multiple sets of training data using one-dimensional word sequences;

[0066] Step 7) Train the neural network model using the training data to obtain the power sequence prediction model;

[0067] Step 8) Obtain the current input data and convert it into a word encoding sequence. Input the word encoding sequence into the power sequence prediction model to obtain the power sequence prediction result.

[0068] Example 2:

[0069] A semantic source-grid-load-storage collaborative power sequence prediction model training method is provided, with the same technical content as in Example 1. Furthermore, the original data for source-grid-load-storage include weather, load, wind power, and photovoltaic power.

[0070] Example 3:

[0071] A semantic source-grid-load-storage collaborative power sequence prediction model training method, the technical content of which is the same as any one of embodiments 1-2, further wherein the original data of source-grid-load-storage includes actual original data and simulated original data;

[0072] The method for generating the simulated original data is as follows:

[0073] Weather data is generated based on historical meteorological data;

[0074] Load power data is generated based on a typical daily load baseline, user behavior correction factors, economic development trend correction factors, and random disturbances, namely:

[0075]

[0076] In the formula, For time period The load power, For hours The corresponding typical daily load benchmark, This refers to the user behavior correction coefficient corresponding to weekdays, weekends, or holidays. This is a seasonal correction factor. This is a correction factor for economic development trends. , and The determination is based on day type, seasonal factors, and economic development factors, respectively. Determined based on load differences between weekdays, weekends, or holidays; Determined based on the impact of monthly, seasonal, or temperature changes on the load; The correction factor is determined based on the annual load growth rate, the level of regional economic development, or historical electricity consumption growth trends. The correction factor can be obtained through historical load data statistics, table lookup, or fitting methods. This is a random disturbance term used to characterize random fluctuations that cannot be directly modeled in load forecasting. The random disturbance term can be a zero-mean random variable, whose variance is determined based on historical load errors or a preset disturbance intensity, and the generated load power is constrained to be no less than zero and not exceeding a preset load upper limit.

[0077] Power output curves are generated using wind power density and solar radiation models;

[0078] The wind power density is shown below:

[0079]

[0080] In the formula, For time period wind power density, air density, For time period The wind speed; wind power output can be generated based on the wind speed segmented output model: when the wind speed is lower than the cut-in wind speed or higher than the cut-out wind speed, the wind power output is... When the wind speed is between the cut-in wind speed and the rated wind speed, the wind power output increases with the increase of wind speed; when the wind speed is between the rated wind speed and the cut-out wind speed, the wind power output is the rated output.

[0081] The solar radiation model is shown below:

[0082]

[0083] In the formula, For time period The intensity of solar radiation, The maximum solar radiation intensity, For the current hour, and These are sunrise and sunset times, respectively. This is a weather correction factor; photovoltaic output can be generated based on solar radiation intensity, photovoltaic installed capacity, and photovoltaic conversion efficiency. Determined based on factors such as weather type, cloud cover, rainfall, haze, or atmospheric transparency, it is used to characterize the attenuation effect of weather conditions on solar radiation intensity. The range of values ​​is The value is close to 1 in sunny weather, and decreases accordingly in cloudy, rainy, or low-visibility weather. The specific value can be determined by statistical analysis of historical radiation data or by referring to tables based on weather type.

[0084] The energy storage state is generated according to the charging and discharging strategy, that is:

[0085]

[0086] In the formula, For time period The state of charge of energy storage, and Time periods The charging power and discharging power, and These are charging efficiency and discharging efficiency, respectively. For the rated capacity of energy storage, For time step.

[0087] Example 4:

[0088] A semantically-based training method for a source-grid-load-storage coordinated power sequence prediction model, with technical content identical to any one of embodiments 1-3, further comprising normalizing the original source-grid-load-storage data so that the numerical range of the original source-grid-load-storage data is 0. ~ 1.

[0089] Example 5:

[0090] A semantic source-grid-load-storage collaborative power sequence prediction model training method, with the same technical content as any one of embodiments 1-4, further wherein the word encoding includes data type, interval number, and trend; the trend is the change trend of the data relative to the previous moment, including rising, falling, and stable.

[0091] Example 6:

[0092] A semantic source-grid-load-storage collaborative power sequence prediction model training method, with the same technical content as any one of embodiments 1-5, further wherein in step 4), the load, wind power, photovoltaic, and weather words at the same time are arranged in a fixed order in the one-dimensional word sequence.

[0093] Example 7:

[0094] A semantic source-grid-load-storage collaborative power sequence prediction model training method, with the same technical content as any one of embodiments 1-6, further wherein the one-dimensional word sequence also carries timestamp words.

[0095] Example 8:

[0096] A semantic source-grid-load-storage collaborative power sequence prediction model training method, with the same technical content as any one of embodiments 1-7, further wherein the input data type of the power sequence prediction model is the original data of source-grid-load-storage other data besides the output data.

[0097] Example 9:

[0098] A semantic source-grid-load-storage coordinated power sequence prediction model training method, with technical content identical to any one of embodiments 1-8, further comprising, in step 6), the step of constructing multiple sets of training data using a one-dimensional word sequence includes:

[0099] Step 6.1) Use the sliding window technique to extract consecutive words from the one-dimensional word sequence; the length of the sliding window is determined based on the data sampling interval, the target prediction step size, the temporal correlation between variables, and the prediction error of the validation set.

[0100] Step 6.2) Randomly hide one word of the same data type as the output word in a series of consecutive words using a masking mechanism, and use the hidden word as the output and the other unhidden words as the input;

[0101] Step 6.3) Repeat steps 6.1)-6.2) until multiple sets of training data are constructed.

[0102] Example 10:

[0103] A semantic source-grid-load-storage coordinated power sequence prediction model training method, with technical content identical to any one of embodiments 1-9, further wherein the training objective of the neural network model is to minimize the prediction error, the loss function is the cross-entropy loss function, and the optimization algorithm is Adam or SGD. Figure 1 As shown, the neural network model includes an embedding lookup layer, a context average aggregation layer, a three-layer fully connected feature extraction layer, a residual connection layer, and a softmax output layer.

[0104] Example 11:

[0105] A semantically-based training method for a source-grid-load-storage coordinated power sequence prediction model is presented below:

[0106] This embodiment transforms power system data into natural language sequences using semantic encoding and sequence modeling techniques, and then utilizes a deep learning model for efficient training. This method not only solves the problems in the background technology but also achieves efficient data utilization and improves the model's generalization ability. The implementation of this invention is described in detail below with reference to the accompanying drawings.

[0107] 1 System Hardware Architecture

[0108] This invention does not rely on specific hardware but can run on general-purpose computing devices, including processors, memory, and input / output interfaces. Optionally, it can be deployed on cloud computing platforms or edge computing nodes to support large-scale data training and real-time inference. The hardware architecture includes:

[0109] Data acquisition module: Used to acquire raw data from power generation, grid, load, and energy storage, including weather, load, wind power, and photovoltaic data. Data can come from sensors, historical databases, or be generated through simulation.

[0110] Processing module: performs semantic encoding, sequence construction, and model training.

[0111] Storage module: Stores vocabulary, training corpus, and model parameters.

[0112] Output module: Used for model inference and scheduling decision support.

[0113] 2. Data Generation and Acquisition

[0114] This invention first generates simulated source-grid-load-storage data to supplement the deficiencies of actual data. Specifically, it generates data for 365 days and 24 time points, covering various operational scenarios (such as seasonal variations and weather fluctuations). The data generation method is based on physical models and statistical distributions, for example:

[0115] Weather data: Parameters such as temperature, humidity, and wind speed are generated based on historical meteorological data.

[0116] Load data: Daily load curves are generated based on user behavior patterns and economic development trends.

[0117] Wind / PV data: Power output curves are generated using wind power density and solar radiation models.

[0118] Energy storage data: Energy storage status is generated based on the charging and discharging strategy.

[0119] By using simulated data, this invention enhances the diversity and representativeness of training data, thus solving the problem of scarce real-world data.

[0120] 3. Semantic coding

[0121] The numerical data is divided into predefined intervals, and independent vocabulary lists are constructed. Specific steps include:

[0122] Interval partitioning: For each variable (such as load, wind power, photovoltaic), its normalized numerical range (0~100%) is divided into multiple intervals (e.g., 50 intervals). The interval partitioning can be adjusted according to the data distribution, for example, by using equal-width partitioning or equal-frequency partitioning.

[0123] Encoding mapping: Mapping each numerical value to a corresponding interval word, taking into account trend information (such as rising, falling, or stable). For example:

[0124] The load value is 0.85, which is in the rising phase. Within the interval, belonging to the 17th interval, it is mapped to a word. .

[0125] Wind power value 0.10, in Within the interval, belonging to the second interval, it is mapped to a word. .

[0126] The photovoltaic value is 0.95. Within the interval, belonging to the 19th interval, it is mapped to a word. .

[0127] This encoding method discretizes continuous numerical values ​​into semantic words, reducing data dimensionality and introducing physical meaning (such as trends) to facilitate model understanding.

[0128] 4. Construct an alternating sequence of energy sentences

[0129] The words for load, wind power, and photovoltaic power at the same time are arranged in a fixed order and interleaved to form a long one-dimensional word sequence. For example:

[0130] Original sequence:

[0131] This sequence structure preserves the coupling information between multiple variables within a given time step (such as the complementary relationship between load and wind power), as well as the evolution information along the time axis (such as load growth trends). The sequence length is determined by the number of time steps, and timestamp words (such as 'T_08' representing 8 o'clock) can optionally be added to enhance the time series information.

[0132] 5. Constructing the training corpus

[0133] Training samples are extracted from interleaved sequences using a sliding window technique. A word-level context prediction task is employed, which involves predicting the missing word in the middle based on the preceding and following N words. Specifically:

[0134] Set the sliding window size (e.g., window length of 5) to extract consecutive words from the sequence as the input sequence.

[0135] By randomly hiding intermediate words using a masking mechanism, context-target pairs are formed. For example:

[0136] Input sequence:

[0137] Output sample:

[0138] Context: (Utilizing simultaneous load and photovoltaic power), Target:'W_02'

[0139] Context: (Utilizing wind power from the previous moment and load from the next moment), Target:

[0140] This training method enables the model to learn the spatiotemporal relationships between multiple variables without the need for additional labeled data, thus achieving self-supervised learning.

[0141] 6. Training a deep semantic model

[0142] A semantic model is used for training. The model input is a sequence of context words, which are converted into vectors through an embedding layer, then processed by a multi-layer neural network to output the probability distribution of the target word. The training objective is to minimize the prediction error using the cross-entropy loss function. The optimization algorithm employs Adam or SGD, with the learning rate dynamically adjusted. During training, the model learns the semantic patterns of the power system; for example, wind power output may decrease when load increases, and peak photovoltaic output corresponds to specific load patterns. Through a large number of training samples, the model captures the inherent patterns of the power system.

[0143] 7. Model Inference and Validation

[0144] After training, the model can be used for power system dispatching tasks. During inference, given partial contextual terms (such as the current load and photovoltaic terms), the model predicts missing variables (such as wind power terms), and then decodes them into numerical ranges to assist in dispatching decisions. The validation phase uses test data to evaluate model performance, with metrics including:

[0145] Prediction accuracy: The accuracy with which the target word is predicted.

[0146] Scheduling efficiency: A comparison between the model-predicted scheduling scheme and the actual optimal scheme.

[0147] Computational latency: The time consumed during the inference process.

[0148] Experiments show that the method of this invention is stable in a variety of scenarios and outperforms traditional numerical methods.

Claims

1. A semantically-based training method for a source-grid-load-storage coordinated power sequence prediction model, characterized in that, Includes the following steps: Step 1) Obtain the raw data of source, grid, load, and storage; Step 2) Normalize the raw data of source, grid, load and storage, and divide the variable interval according to the normalized value range; Step 3) Determine the trends and values ​​of different types of data in the original data, and generate word codes based on the correspondence between the values ​​and variable ranges; Step 4) Integrate the word codes into a one-dimensional word sequence according to time order; Step 5) Determine the input data type and output data type of the power sequence prediction model; the output data type is either wind power or photovoltaic power. Step 6) Based on the input and output data types, construct multiple sets of training data using one-dimensional word sequences; Step 7) Train the neural network model using the training data to obtain the power sequence prediction model; Step 8) Obtain the current input data and convert it into a word encoding sequence. Input the word encoding sequence into the power sequence prediction model to obtain the power sequence prediction result.

2. The semantic source-grid-load-storage collaborative power sequence prediction model training method according to claim 1, characterized in that, The raw data for the power generation, grid, load, and energy storage system includes weather, load, wind power, and solar power.

3. The semantic source-grid-load-storage collaborative power sequence prediction model training method according to claim 2, characterized in that, The raw data of source-grid-load-storage includes actual raw data and simulated raw data; The method for generating the simulated original data is as follows: Weather data is generated based on historical meteorological data; Load power data is generated based on a typical daily load baseline, user behavior correction factors, economic development trend correction factors, and random disturbances, namely: In the formula, For time period The load power, For hours The corresponding typical daily load benchmark, This refers to the user behavior correction coefficient corresponding to weekdays, weekends, or holidays. This is a seasonal correction factor. This is a correction factor for economic development trends. For random disturbance terms; Power output curves are generated using wind power density and solar radiation models; The wind power density is shown below: In the formula, For time period wind power density, air density, For time period wind speed; The solar radiation model is shown below: In the formula, For time period The intensity of solar radiation, The maximum solar radiation intensity, For the current hour, and These represent sunrise and sunset times, This is a weather correction factor; The energy storage state is generated according to the charging and discharging strategy, that is: In the formula, For time period The state of charge of energy storage, and Time periods The charging power and discharging power, and These are charging efficiency and discharging efficiency, respectively. For the rated capacity of energy storage, For time step.

4. The semantic source-grid-load-storage collaborative power sequence prediction model training method according to claim 1, characterized in that, After normalization, the raw data of source-grid-load-storage are in the range of 0. ~ 1.

5. The semantic source-grid-load-storage collaborative power sequence prediction model training method according to claim 1, characterized in that, Word encoding includes data type, interval index, and trend; The trend refers to the change in data relative to the previous moment, including rising, falling, and remaining stable.

6. The semantic source-grid-load-storage collaborative power sequence prediction model training method according to claim 1, characterized in that, In the one-dimensional word sequence of step 4), the words for load, wind power, photovoltaic, and weather at the same time are arranged in a fixed order.

7. The semantic source-grid-load-storage collaborative power sequence prediction model training method according to claim 6, characterized in that, The one-dimensional word sequence also carries timestamped words.

8. The semantic source-grid-load-storage collaborative power sequence prediction model training method according to claim 1, characterized in that, The input data for the power sequence prediction model consists of the original data from the power generation, grid, load, and storage systems, excluding the output data.

9. The semantic source-grid-load-storage collaborative power sequence prediction model training method according to claim 1, characterized in that, Step 6), the steps for constructing multiple sets of training data using one-dimensional word sequences, include: Step 6.1) Extract consecutive words from a one-dimensional word sequence using the sliding window technique; Step 6.2) Randomly hide one word of the same data type as the output word in a series of consecutive words using a masking mechanism, and use the hidden word as the output and the other unhidden words as the input; Step 6.3) Repeat steps 6.1)-6.2) until multiple sets of training data are constructed.

10. The semantic source-grid-load-storage collaborative power sequence prediction model training method according to claim 1, characterized in that, The training objective of a neural network model is to minimize the prediction error. The loss function is the cross-entropy loss function, and the optimization algorithm is Adam or SGD.

Images