A Deep Learning-Based Method and System for Cooperative Pre-planning and Real-time Optimization Scheduling of Power Grids

Abstract

This invention discloses a deep learning-based method and system for collaborative pre-planning and real-time optimized scheduling of power grids, belonging to the field of power system dispatching technology. The method first fuses multi-source heterogeneous data, extracts textual semantic features using natural language processing techniques, and combines these features with numerical features to form a unified feature representation. Next, it uses a deep neural network for day-ahead collaborative pre-planning to generate day-ahead dispatching schemes that meet operational constraints. Finally, based on real-time system status and ultra-short-term forecast information, a rapid optimization model is constructed to dynamically adjust the day-ahead plan to address the uncertainties in renewable energy output and load forecasting. This invention achieves deep fusion of multi-source information and collaborative optimization of dispatching decisions, improving the economy, adaptability, and operational reliability of power grid dispatching.

Description

Technical Field

[0001] This invention relates to the field of power dispatching, and in particular to a method and system for collaborative pre-planning and real-time optimization dispatching of power grids based on deep learning. Background Technology

[0002] The power system is the cornerstone of modern society, providing a stable and reliable power supply to all sectors and supporting the normal operation of daily life, industrial production, and socio-economic activities. Power grid dispatch is the core guarantee for the safe and stable operation of the power system. Through real-time monitoring and optimized allocation of generation and consumption loads, it ensures a dynamic balance between power supply and demand, maintains voltage and frequency stability, reduces generation costs, improves energy efficiency, and provides rapid response in the event of sudden faults or natural disasters, minimizing losses and providing reliable power support for the sustainable development of the economy and society. Therefore, accurate and efficient power grid dispatch is crucial to the power system.

[0003] With the continued advancement of high-proportion renewable energy integration and the intelligent transformation of the power system, the fundamental challenge facing power grid dispatch lies in how to efficiently integrate multi-source heterogeneous data and, based on this, cope with significantly increased uncertainty. The core of traditional dispatch systems relies on structured numerical data and optimization algorithms based on deterministic models. While these methods possess stability, their limitations are becoming increasingly apparent: On the one hand, the large amount of unstructured data generated during dispatch, such as text reports and operation logs, contains valuable expert experience and contextual logic, which current technologies struggle to deeply understand and effectively integrate, resulting in incomplete decision-making information. On the other hand, day-ahead planning and real-time operational adjustments are typically treated as relatively independent phases. Day-ahead planning is mostly generated based on day-ahead forecast data and lacks flexibility once formulated; while real-time adjustments rely heavily on local controllers or simplified rule responses, which, while ensuring safety, struggle to achieve both rapid action and overall economic optimization. This disconnect between "planning" and "adjustment" often results in high adjustment costs and insufficient overall coordination when the system responds to real-time fluctuations in renewable energy and load, making it difficult to achieve economically optimal dispatch across all time periods under safety constraints. Summary of the Invention

[0004] To address the above problems, this invention provides a deep learning-based method for collaborative pre-planning and real-time optimization scheduling of power grids. This method can effectively integrate multi-source heterogeneous data, deeply mine textual semantic information, and realize intelligent day-ahead planning and real-time optimization of power grid scheduling, thereby improving the accuracy, economy, and adaptability of power grid scheduling.

[0005] The technical solution of this invention is: a deep learning-based method for collaborative pre-planning and real-time optimization scheduling of power grids, comprising the following steps:

[0006] Multi-source data preprocessing and fusion to form feature vectors;

[0007] The feature vectors are input into a deep neural network for training, with the goal of minimizing scheduling costs, and the day-ahead plan is output.

[0008] Based on the day-ahead plan, a real-time optimization model is constructed with the goal of minimizing adjustment costs, using real-time system status and ultra-short-term forecast information. This model solves for the real-time optimized output command that satisfies power balance, unit output, and ramp-up constraints, enabling online dynamic correction of the day-ahead plan.

[0009] Multi-source data includes historical load data, wind and solar power output data, unit operating parameters, market environment data, and meteorological data. Among them, unit operating parameters include minimum technical output, maximum output, ramp rate, start-up and shutdown time, and cost coefficient; market environment data includes nodal marginal electricity price, ancillary service market price, and fuel price; and meteorological data includes temperature, humidity, wind speed, and irradiance.

[0010] Multi-source data preprocessing includes:

[0011] For structured data, preprocessing operations including missing value imputation, outlier detection and correction, z-score normalization or max-min normalization are used to obtain standardized numerical features.

[0012] For unstructured data, a natural language processing model is used to perform deep feature extraction on power grid dispatch text reports to obtain text semantic features;

[0013] By deeply fusing textual semantic features with numerical features, feature vectors are generated.

[0014] For power grid dispatch text reports, firstly, the Word2Vec model is used to initialize word vectors in the text data of the report, and a professional power field word vector space is constructed.

[0015] Secondly, a BiLSTM-CRF model is used for named entity recognition. The BiLSTM layer captures long-term dependencies in the text through a gating mechanism, preserving important scheduling operation information; the CRF layer performs global optimization on the output label sequence.

[0016] Then, the BERT model is used to extract relationships, transforming the device associations and operational logic described in the text into structured relationship features;

[0017] Finally, the entities and relations extracted by the RoBERTa model are used for deep semantic encoding to generate high-quality vector representations. The extracted entities, relations, and context are the inputs to the RoBERTa model. The RoBERTa model uses a multi-layer Transformer architecture and a self-attention mechanism to encode the input sequence and generate a context-related vector representation for each character.

[0018] For entities identified from power grid dispatch text reports, an attention-weighted encoding strategy is used to generate entity vectors. Its formula is:

[0019]

[0020] Where L is the number of characters contained in the entity name; This represents the hidden state vector of the j-th character output from the last layer of the RoBERTa model; The attention weight represents the j-th character;

[0021] For relational encoding The fusion coding method based on multilayer perceptron is adopted, and its formula is:

[0022]

[0023] in, It is the hidden state marked by [CLS] at the beginning of the sequence; and These are the vector representations of the head and tail entities, respectively; MLP is a multilayer perceptron.

[0024] All entity and relation vectors are aggregated using an average pooling strategy to form a document-level semantic feature representation. :

[0025]

[0026] Where N represents the total number of entities identified in a single scheduling document; M represents the total number of relations extracted from a single scheduling document;

[0027] This represents the i-th entity vector in the document; This represents the j-th relation vector in the document.

[0028] The deep neural network includes an input layer, multiple hidden layers, and an output layer. The hidden layers use the ReLU activation function. The output layer employs an activation function design, specifically: the sigmoid function combined with the Gumbel-Softmax technique is used for the unit start-up and shutdown states; and the unit output is scaled using the Tanh function with upper and lower bound constraints.

[0029] During network training, the network takes feature vectors as input and minimizes the scheduling cost evaluation function. The target value is , and the gradient descent algorithm is used for end-to-end training to learn the mapping from system state to optimal scheduling policy.

[0030] The formula for the scheduling cost evaluation function is as follows:

[0031]

[0032] Where T is the total number of scheduling periods; t is the index of the scheduling period; and i is the generator set index, belonging to the set. ; and These represent the actual active power output monitoring value and the actual start-stop status monitoring value of unit i during time period t, respectively; and These represent the preset operating cost coefficients of unit i itself; This represents the total expected power output of the generating unit during time period t; This represents the total actual power output of the generating unit during time period t; This represents the preset energy curtailment penalty weighting coefficient.

[0033] Based on the current plan, and using real-time data on the total system load, actual total output of new energy sources, and actual output of each unit, the overall deviation between the current system's net load demand and the planned output is calculated. If a power imbalance is detected, a real-time optimization model is constructed and solved to minimize adjustment costs, as follows:

[0034]

[0035] in, Units awaiting demand In the upcoming execution period Optimized output value; For the same unit corresponding to the same time period in the original plan. The original planned output value; For the unit The unit adjustment cost coefficient; This refers to the collection of traditional generator sets authorized by the dispatching system to participate in this real-time power adjustment.

[0036] The power balance constraint is:

[0037]

[0038] in, This represents the sum of the optimized planned output of all units involved in the adjustment; This represents the predicted total load of the system in time period t. This represents the total actual output of the unit during the current time period t.

[0039] The unit output constraint is:

[0040]

[0041] in, For the unit During the execution period Optimized output value; and They refer to the generating units respectively. The minimum and maximum active power that can be continuously output;

[0042] The unit's ramp-up constraint is:

[0043]

[0044] in, For the unit During the execution period Optimized output value; For the unit In the previous execution period Optimized output value; This refers to the maximum ramp rate of unit i.

[0045] A deep learning-based power grid collaborative pre-planning and real-time optimization scheduling system includes:

[0046] The fusion module is used for preprocessing and fusing multi-source data to form feature vectors.

[0047] The output module is used to input feature vectors into a deep neural network for training, with the goal of minimizing scheduling costs, and output the daily schedule.

[0048] The correction module is used to construct a real-time optimization model based on the day-ahead plan and real-time system status and ultra-short-term forecast information, with the goal of minimizing adjustment costs. It solves for the real-time optimized output command that satisfies power balance, unit output and ramp-up constraints, thereby realizing online dynamic correction of the day-ahead plan.

[0049] The beneficial effects of this invention are as follows: by organically combining deep fusion and feature extraction of multi-source data, day-ahead collaborative pre-planning based on deep learning, and real-time optimization and adjustment, the intelligence level and economy of power grid dispatch are significantly improved; it not only realizes efficient fusion and semantic understanding of multi-dimensional information such as text and numerical data, but also constructs a day-ahead pre-planning and real-time optimization collaborative decision-making mechanism, which can dynamically respond to new energy fluctuations and load changes while ensuring system safety and stability, effectively reduce overall dispatch costs, and improve the economy and adaptability of power grid operation. Attached Figure Description

[0050] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0051] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0052] like Figure 1 As shown, the deep learning-based power grid collaborative pre-planning and real-time optimization scheduling method of the present invention includes the following steps:

[0053] Multi-source data preprocessing and fusion: By preprocessing and imputing missing values ​​of structured data such as load, unit, market, and meteorology, and combining natural language processing models to perform deep semantic parsing of power grid dispatch text reports, entity and relation features are extracted. Finally, the text semantic features and numerical features are deeply fused to generate a unified high-dimensional feature vector.

[0054] Day-ahead collaborative pre-planning decision-making: The fused feature vectors are input into a deep neural network for training. The network adopts a specific activation function and a differentiable discrete decision design to minimize scheduling costs. It learns end-to-end and outputs day-ahead unit start-up, shutdown and output plans that meet safety and economic constraints.

[0055] Real-time dynamic optimization and adjustment: Based on the day-ahead plan, a real-time optimization model is constructed with the goal of minimizing adjustment costs, based on real-time system status and ultra-short-term forecast information. The model is then quickly solved to obtain the real-time optimized output command that satisfies power balance, unit output and ramp-up constraints, thereby realizing online dynamic correction of the day-ahead plan.

[0056] Specifically,

[0057] Step 1: Integrate multi-source data such as load, power forecast, and unit parameters through deep learning models, and use various models to preprocess and complete the data in order to extract feature vectors that characterize the future operation of the power grid.

[0058] This invention first performs multi-source data fusion and high-dimensional feature extraction. It first acquires high-dimensional heterogeneous data from different sources, including historical load data, wind and solar power output data, unit operating parameters, market environment data, and meteorological data. Specifically, historical load data includes 15-minute load data from the past 3-5 years, including different patterns such as weekdays, weekends, and holidays; wind and solar power output data includes historical actual output and forecast data, with forecast data including day-ahead forecasts and ultra-short-term forecasts (specific times can be set according to actual conditions); unit operating parameters include minimum technical output, maximum output, ramp rate, start-up and shutdown time, and cost coefficients; market environment data includes nodal marginal electricity prices, ancillary service market prices, and fuel prices; and meteorological data includes meteorological factors affecting load and renewable energy output, such as temperature, humidity, wind speed, and irradiance.

[0059] Secondly, the above data undergoes preprocessing and completion. For structured data, this invention performs preprocessing operations including missing value imputation, outlier detection and correction, and z-score normalization or max-min normalization to obtain standardized numerical features. For unstructured data, this invention employs a natural language processing model for deep feature extraction. Specifically, for a power grid dispatch text report, this invention first uses the Word2Vec model to initialize word vectors in the text data of the report, constructing a professional power domain word vector space to effectively capture the semantic information of power professional terms. Based on this, a BiLSTM-CRF model is used for named entity recognition. The BiLSTM layer captures long-term dependencies in the text through its unique gating mechanism, preserving important dispatch operation information, while the CRF layer globally optimizes the output label sequence to ensure the accuracy of entity recognition. Furthermore, a BERT model is used for relation extraction, transforming the equipment associations and operational logic described in the text into structured relation features. Then, the entities and relations extracted by the RoBERTa model are used for deep semantic encoding to generate high-quality vector representations. The extracted entities, relations, and context serve as inputs to the RoBERTa model. The model employs a multi-layer Transformer architecture and a self-attention mechanism to encode the input sequence, generating a context-dependent vector representation for each character. For entities identified from power grid dispatch text reports, an attention-weighted encoding strategy is used to generate entity vectors. Its formula is:

[0060]

[0061] Where L is the number of characters contained in the entity name; This represents the hidden state vector of the j-th character output from the last layer of the RoBERTa model, which carries the deep semantic information of the character in the specific power grid scheduling context. The attention weight representing the j-th character can be calculated using TF-IDF weights, reflecting the importance of different characters in the semantic composition of the entity.

[0062] For relational encoding The fusion coding method based on multilayer perceptron is adopted, and its formula is:

[0063]

[0064] in, It is the hidden state marked by [CLS] at the beginning of the sequence, which carries the semantic information of the entire sequence; and These are the vector representations of the head entity and the tail entity, respectively; MLP is a multilayer perceptron used to fuse semantic information from different sources.

[0065] Finally, this invention uses an average pooling aggregation strategy to aggregate all entity and relation vectors, forming a document-level semantic feature representation. :

[0066]

[0067] Where N represents the total number of entities identified in a single scheduling document; M represents the total number of relations extracted from a single scheduling document;

[0068] This represents the i-th entity vector in the document; This represents the j-th relation vector in the document. The Word2Vec, BiLSTM-CRF, BERT, and RoBERTa models described above constitute a collaborative processing pipeline from lexical semantic initialization to structured entity and relation recognition, and finally deep contextual semantic encoding. Through this process, the RoBERTa model can transform unstructured power grid dispatch text reports into vector representations rich in semantic information. These vectors not only preserve the original text semantics but also incorporate professional knowledge from the power grid dispatch domain, providing high-quality semantic feature input for subsequent intelligent dispatch decisions.

[0069] Finally, this invention concatenates and fuses the preprocessed numerical feature matrix and document-level semantic feature vector at a self-attention layer, using them together as input to the subsequent feature extraction network. This achieves complementarity and enhancement between textual semantic information and numerical statistical information. The feature extraction network employs a hierarchical architecture with clear division of labor: the bottom-layer one-dimensional convolutional neural network is responsible for extracting local temporal patterns and short-term correlations in the numerical sequence; the middle-layer bidirectional long short-term memory network is used to model long-term dependencies and periodic patterns in the data; the top-layer self-attention mechanism dynamically evaluates and focuses on features at the most critical time points for scheduling decisions; and finally, a fully connected layer projects the fused features onto a fixed-dimensional dense vector S. f Preferably, S f The dimensionality is controlled between 128 and 512, balancing information integrity and model efficiency.

[0070] Step 2, Deep Learning-Based Collaborative Pre-planning for the Same Day Network: First, this invention uses the feature vector S generated in Step 1... fThe input is fed into a day-ahead policy network composed of a deep neural network and trained. The network structure includes an input layer, multiple hidden layers, and an output layer. The hidden layers use the ReLU activation function, and the output layer employs a specific activation function design: the sigmoid function combined with the Gumbel-Softmax technique is used for differentiable discrete decisions regarding unit start-up and shutdown states; unit output is scaled using a Tanh function with upper and lower bound constraints to ensure that the output meets physical constraints. This design ensures that the output meets both physical and operational constraints and enables end-to-end training of the entire network. During network training, the network uses the feature vector S from historical data... f As input, minimize the scheduling cost evaluation function. The objective is to learn the mapping from system state to optimal scheduling policy using gradient descent for end-to-end training. The scheduling cost evaluation function, i.e., the objective function, is as follows:

[0071]

[0072] Where T is the total number of scheduling periods, which can be 96, corresponding to 96 15-minute intervals in a day; t is the scheduling period index, ranging from 1 to T; and i is the generator set index, belonging to the set... ; and These represent the actual active power output monitoring value and the actual start-stop status monitoring value of unit i during time period t, respectively; and These represent the preset operating cost coefficients of unit i itself; This represents the total expected power output of the generating unit during time period t; This represents the total actual power output of the generating unit during time period t; This represents the preset energy curtailment penalty weighting coefficient.

[0073] Based on the above objective function, the network learns and ultimately outputs a complete and executable day-ahead scheduling decision scheme, which forms the operating benchmark for each time period of the next day.

[0074] Step 3: Optimization of Day-ahead Dispatch Decision Scheme: Based on the day-ahead dispatch scheme formed in Step 2, and according to the latest real-time operating status of the power grid and ultra-short-term forecast information, this invention performs online, rapid, and economical optimal correction of the plan to cope with uncertainties such as new energy fluctuations and load forecasting errors.

[0075] First, based on real-time data collection of the system's total load, actual total output of new energy sources, and actual output of each unit, this invention calculates the overall deviation between the current system's net load demand and planned output. If a power imbalance is detected, a real-time optimization model is immediately constructed and solved. This model aims to minimize adjustment costs and is formalized as follows:

[0076]

[0077] in, Units awaiting demand In the upcoming execution period The optimized output value is the decision variable of the model; For the same unit corresponding to the same time period in the original plan. The original planned output value; For the unit The unit regulation cost coefficient is a parameter attribute inherent to the unit, reflecting the additional costs such as fuel and losses corresponding to the unit's output regulation service when providing output regulation services. This refers to the collection of traditional generator sets authorized by the dispatching system to participate in this real-time power adjustment.

[0078] Meanwhile, the optimization process must adhere to global constraints, including real-time power balance constraints, upper and lower limits of unit output constraints, and unit ramp-up rate constraints. Real-time power balance constraints are a fundamental condition for ensuring power system frequency stability, and the formula is as follows:

[0079]

[0080] in, This represents the sum of the optimized planned output of all units involved in the adjustment, and its value is determined by the objective function optimization. The total load forecast value of the system in time period t is the standard output of the standard configuration system of the power dispatch center, which is used to characterize the expected power demand of the system in future time periods. This represents the total actual output of the unit during the current time period t.

[0081] The upper and lower limits of unit output are constrained to ensure that the optimized unit output value does not exceed the unit's own safe operating range. The formula is as follows:

[0082]

[0083] in, For the unit During the execution period Optimized output value; and They refer to the generating units respectively. The minimum and maximum active power that can be continuously output under safe and stable operating conditions are the inherent technical parameters of the generating unit.

[0084] The unit ramp-up rate constraint ensures that the unit output changes smoothly, avoiding equipment damage, operational instability, or grid frequency fluctuations caused by sudden increases / decreases in power. The formula is as follows:

[0085]

[0086] in, For the unit During the execution period Optimized output value; For the unit In the previous execution period Optimized output value; The maximum ramp rate of unit i refers to the maximum permissible change in active power that the unit can increase or decrease between two adjacent scheduling periods. It is an inherent technical parameter of the unit.

[0087] Finally, since the objective function and constraint formulas described above are all linear programming problems, efficient algorithms such as the simplex method or interior-point method can be used for rapid solution. The results satisfy both the objective function and power balance constraints, and this solution set constitutes the latest scheduling strategy after real-time optimization. This strategy, after being formatted into standard scheduling instructions, can be issued to relevant power plants for execution.

[0088] A deep learning-based power grid collaborative pre-planning and real-time optimization scheduling system includes:

[0089] The fusion module is used for preprocessing and fusing multi-source data to form feature vectors.

[0090] The output module is used to input feature vectors into a deep neural network for training, with the goal of minimizing scheduling costs, and output the daily schedule.

[0091] The correction module is used to construct a real-time optimization model based on the day-ahead plan and real-time system status and ultra-short-term forecast information, with the goal of minimizing adjustment costs. It solves for the real-time optimized output command that satisfies power balance, unit output and ramp-up constraints, thereby realizing online dynamic correction of the day-ahead plan.

[0092] This invention enables intelligent day-ahead planning and real-time optimization of power grid dispatch, improving the accuracy, economy, and adaptability of power grid dispatch, and providing technical support for the safe and stable operation of new power systems.

[0093] The above specific embodiments have further described in detail the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A deep learning-based method for collaborative pre-planning and real-time optimized scheduling of power grids, characterized in that, Includes the following steps: Multi-source data preprocessing and fusion to form feature vectors; The feature vectors are input into a deep neural network for training, with the goal of minimizing scheduling costs, and the day-ahead plan is output. Based on the day-ahead plan, a real-time optimization model is constructed with the goal of minimizing adjustment costs, using real-time system status and ultra-short-term forecast information. This model solves for the real-time optimized output command that satisfies power balance, unit output, and ramp-up constraints, enabling online dynamic correction of the day-ahead plan.

2. The deep learning-based power grid collaborative pre-planning and real-time optimization scheduling method according to claim 1, characterized in that: Multi-source data includes historical load data, wind and solar power output data, unit operating parameters, market environment data, and meteorological data. Among them, unit operating parameters include minimum technical output, maximum output, ramp rate, start-up and shutdown time, and cost coefficient; market environment data includes nodal marginal electricity price, ancillary service market price, and fuel price; and meteorological data includes temperature, humidity, wind speed, and irradiance.

3. The deep learning-based power grid collaborative pre-planning and real-time optimization scheduling method according to claim 1, characterized in that: Multi-source data preprocessing includes: For structured data, preprocessing operations including missing value imputation, outlier detection and correction, z-score normalization or max-min normalization are used to obtain standardized numerical features. For unstructured data, a natural language processing model is used to perform deep feature extraction on power grid dispatch text reports to obtain text semantic features; By deeply fusing textual semantic features with numerical features, feature vectors are generated.

4. The deep learning-based power grid collaborative pre-planning and real-time optimization scheduling method according to claim 3, characterized in that: For power grid dispatch text reports, firstly, the Word2Vec model is used to initialize word vectors in the text data of the report, and a professional power field word vector space is constructed. Secondly, a BiLSTM-CRF model is used for named entity recognition. The BiLSTM layer captures long-term dependencies in the text through a gating mechanism, preserving important scheduling operation information; the CRF layer performs global optimization on the output label sequence. Then, the BERT model is used to extract relationships, transforming the device associations and operational logic described in the text into structured relationship features; Finally, the entities and relations extracted by the RoBERTa model are used for deep semantic encoding to generate high-quality vector representations. The extracted entities, relations, and context are the inputs to the RoBERTa model. The RoBERTa model uses a multi-layer Transformer architecture and a self-attention mechanism to encode the input sequence and generate a context-related vector representation for each character.

5. The deep learning-based power grid collaborative pre-planning and real-time optimization scheduling method according to claim 4, characterized in that: For entities identified from power grid dispatch text reports, an attention-weighted encoding strategy is used to generate entity vectors. Its formula is: Where L is the number of characters contained in the entity name; This represents the hidden state vector of the j-th character output from the last layer of the RoBERTa model; The attention weight represents the j-th character; For relational encoding The fusion coding method based on multilayer perceptron is adopted, and its formula is: in, It is the hidden state marked by [CLS] at the beginning of the sequence; and These are the vector representations of the head and tail entities, respectively; MLP is a multilayer perceptron. All entity and relation vectors are aggregated using an average pooling strategy to form a document-level semantic feature representation. : Where N represents the total number of entities identified in a single scheduling document; M represents the total number of relations extracted from a single scheduling document; This represents the i-th entity vector in the document; This represents the j-th relation vector in the document.

6. The deep learning-based power grid collaborative pre-planning and real-time optimization scheduling method according to claim 1, characterized in that: The deep neural network includes an input layer, multiple hidden layers, and an output layer. The hidden layers use the ReLU activation function. The output layer employs an activation function design, specifically: the sigmoid function combined with the Gumbel-Softmax technique is used for the unit start-up and shutdown states; and the unit output is scaled using the Tanh function with upper and lower bound constraints. During network training, the network takes feature vectors as input and minimizes the scheduling cost evaluation function. The target value is , and the gradient descent algorithm is used for end-to-end training to learn the mapping from system state to optimal scheduling policy.

7. The deep learning-based power grid collaborative pre-planning and real-time optimization scheduling method according to claim 6, characterized in that: The formula for the scheduling cost evaluation function is as follows: Where T is the total number of scheduling periods; t is the index of the scheduling period; and i is the generator set index, belonging to the set. ; and These represent the actual active power output monitoring value and the actual start-stop status monitoring value of unit i during time period t, respectively; and These represent the preset operating cost coefficients of unit i itself; This represents the total expected power output of the generating unit during time period t; This represents the total actual power output of the generating unit during time period t; This represents the preset energy curtailment penalty weighting coefficient.

8. The deep learning-based power grid collaborative pre-planning and real-time optimization scheduling method according to claim 1, characterized in that: Based on the current plan, and using real-time data on the total system load, actual total output of new energy sources, and actual output of each unit, the overall deviation between the current system's net load demand and the planned output is calculated. If a power imbalance is detected, a real-time optimization model is constructed and solved to minimize adjustment costs, as follows: in, Units awaiting demand In the upcoming execution period Optimized output value; For the same unit corresponding to the same time period in the original plan. The original planned output value; For the unit The unit adjustment cost coefficient; This refers to the collection of traditional generator sets authorized by the dispatching system to participate in this real-time power adjustment.

9. The deep learning-based power grid collaborative pre-planning and real-time optimization scheduling method according to claim 8, characterized in that: The power balance constraint is: in, This represents the sum of the optimized planned output of all units involved in the adjustment; This represents the predicted total load of the system in time period t. This represents the total actual output of the unit during the current time period t. The unit output constraint is: in, For the unit During the execution period Optimized output value; and They refer to the generating units respectively. The minimum and maximum active power that can be continuously output; The unit's ramp-up constraint is: in, For the unit During the execution period Optimized output value; For the unit In the previous execution period Optimized output value; This refers to the maximum ramp rate of unit i.

10. A deep learning-based power grid collaborative pre-planning and real-time optimization scheduling system, characterized in that, include: The fusion module is used for preprocessing and fusing multi-source data to form feature vectors. The output module is used to input feature vectors into a deep neural network for training, with the goal of minimizing scheduling costs, and output the daily schedule. The correction module is used to construct a real-time optimization model based on the day-ahead plan and real-time system status and ultra-short-term forecast information, with the goal of minimizing adjustment costs. It solves for the real-time optimized output command that satisfies power balance, unit output and ramp-up constraints, thereby realizing online dynamic correction of the day-ahead plan.

Images