Fuel gas and steam combined cycle unit power plant load prediction and optimal scheduling method and system based on artificial intelligence

Abstract

The invention discloses a gas-steam combined cycle unit power plant load prediction and optimal scheduling method and system based on artificial intelligence, and relates to the technical field of artificial intelligence. Load data and unit output of a gas-steam combined cycle unit power plant are recorded in real time; load fluctuation characteristics are extracted based on the load data, prediction error feedback is obtained, an adaptive time window mechanism is introduced, and the time window length of next prediction is adjusted according to the current load fluctuation characteristics and the prediction error feedback; extracting a multi-time scale fusion feature of the load data of each time step in the time window according to the updated time window length; constructing a load prediction model, and predicting future load data based on the multi-time scale fusion features of each time step; based on the current unit output, the future load data and the self-adaptive time window mechanism, constructing a reinforcement learning model to optimize the unit output adjustment value of the gas-steam combined cycle unit power plant; and the self-adaptability of load prediction and optimal scheduling is realized.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and particularly to a method and system for power load prediction and optimal scheduling of a gas-steam combined cycle unit power plant based on artificial intelligence. Background Art

[0002] A combined cycle gas turbine (CCGT) is a highly efficient and clean power generation technology. It combines a gas turbine and a steam turbine, uses the waste heat of the high-temperature flue gas discharged from the gas turbine to generate steam, and drives the steam turbine to generate electricity, thereby greatly improving the power generation efficiency and reducing emissions. The CCGT power plant has good flexibility and fast startup, and is an important power source in the power system.

[0003] Chinese Patent with application publication number CN113107626A discloses a method for predicting the load of a combined cycle power generation unit based on multi-variable LSTM, including the steps of: S1, data collection; S2, data preprocessing; S3, dividing the data of one unit into a training set and a test set, and setting the data of the other unit as a validation set; S4, building a multi-variable input LSTM neural network model; S5, substituting the validation set data into the model for prediction, and using the combined cycle power as the prediction target of the unit load; S6, evaluating the model according to two evaluation indexes, namely the loss value in model training and the root mean square error in prediction; S7, generating a true value-predicted value curve graph.

[0004] The prior art usually adopts a fixed time window and is difficult to adapt to the dynamic characteristics of load changes. Summary of the Invention

[0005] This application aims to solve at least one of the technical problems in the related art to some extent. For this reason, an object of this application is to propose a method and system for power load prediction and optimal scheduling of a gas-steam combined cycle unit power plant based on artificial intelligence, which realizes the comprehensive modeling and prediction of short-term load fluctuations and long-term trends.

[0006] One aspect of this application provides a method for power load prediction and optimal scheduling of a gas-steam combined cycle unit power plant based on artificial intelligence, including: Step S100: Record the load data and unit output of the gas-steam combined cycle unit power plant in real time; Step S200: Extract the load fluctuation characteristics based on the load data, obtain the prediction error feedback, introduce an adaptive time window mechanism, and adjust the time window length of the next prediction according to the current load fluctuation characteristics and prediction error feedback; Step S300: Extract the multi-time scale fusion characteristics of the load data at each time step within the time window according to the updated time window length; Step S400: Construct a load forecasting model to predict future load data based on the multi-time-scale fusion features at each time step; Step S500: Based on the current unit output, future load data, and the adaptive time window mechanism, construct a reinforcement learning model to optimize the unit output adjustment value of the gas-steam combined cycle power plant; The specific method for real-time recording of the load data and unit output of the gas-steam combined cycle power plant is as follows: Real-time record the load data and unit output of the gas-steam combined cycle power plant and store them in the database. Obtain the load data of the most recent TN time steps from the power plant's database , perform data cleaning on the load data to remove abnormal data and invalid data, and interpolate the missing data; where, represents the load data at the t-th time step, and TN is the total number of time steps; The specific method for extracting the load fluctuation characteristics based on the load data, obtaining the prediction error feedback, introducing the adaptive time window mechanism, and adjusting the time window length of the next prediction according to the current load fluctuation characteristics and prediction error feedback is as follows: Step S210: Divide NI types of time window levels according to different time spans. Assume that the time span of the ni-th time window level is , and calculate the number of time windows at this time window level; Step S220: For the ni-th time window level, obtain the load data of the nj-th time window therein ; Step S230: For the nj-th time window in the ni-th time window level, calculate the load fluctuation characteristics of this time window according to the time window length and the load data of adjacent time steps in the corresponding time window ; Step S240: According to the predicted value of the load data and the true value , calculate the prediction error feedback ; Step S250: Preset the load fluctuation threshold and the prediction error fluctuation threshold , compare the load fluctuation characteristics with the load fluctuation threshold, and the prediction error feedback with the prediction error fluctuation threshold, and update the length of the time window for the next prediction based on the adaptive time window mechanism; The specific method for updating the length of the time window for the next prediction based on the adaptive time window mechanism is as follows: When the load fluctuation characteristics are greater than the load fluctuation threshold or the prediction error feedback is greater than the prediction error fluctuation threshold When calculating, obtain the time difference between the length of the time window and the adjustment step of the time window length, and compare the size of the time difference with the minimum value of the time window length. Update the length of the time window for the next prediction to the larger value between the two. When the load fluctuation feature is less than the load fluctuation threshold and the prediction error feedback is less than the prediction error fluctuation threshold, calculate the time sum between the length of the time window and the adjustment step of the time window length, and compare the size of the time sum with the maximum value of the time window length. Update the length of the time window for the next prediction to the smaller value between the two. Otherwise, the length of the time window for the next prediction remains unchanged. Obtain the time difference between them, and compare the time difference with the minimum value of the time window length. Compare the magnitudes between them, and update the length of the time window for the next prediction to the larger value between the two; when the load fluctuation feature is less than the load fluctuation threshold and the prediction error feedback is less than the prediction error fluctuation threshold, calculate the time sum between the length of the time window and the adjustment step of the time window length. Compare the time sum with the maximum value of the time window length. Compare the magnitudes between them, and update the length of the time window for the next prediction to the smaller value between the two; otherwise, the length of the time window for the next prediction remains unchanged; The specific method for extracting the multi-time scale fusion features of the load data at each time step within the time window according to the updated time window length is as follows: Step S310: Based on the updated length of the time window, for each time step, extract the load features belonging to each time window level and the time window corresponding to it. The load features include the load features of the load data in the time domain, frequency domain, and time-frequency domain; Step S320: For each time step, calculate the fusion weight of the load features extracted by the nj-th time window at the ni-th time window level at this time step; ; Step S330: Based on the fusion weight, fuse the load features extracted by different time windows at time step t to obtain the fused load features; ; Step S340: Synthesize the fused load features of different time window levels to obtain the multi-time scale fusion features at the t-th time step; ; The specific method for constructing the load prediction model and predicting future load data based on the multi-time scale fusion features at each time step is as follows: Step S410: Construct a load prediction model, using a long short-term memory neural network model as the basic model and introducing an attention mechanism. The load prediction model includes an input layer, a long short-term memory layer, an attention mechanism layer, an output layer, and a loss function. The loss function includes the mean square error and a regularization term. Optimize and train the load prediction model by minimizing the loss function; Step S420: Take the multi-time scale fusion features as input data and input them into the input layer. Use the long short-term memory layer to learn the temporal features of the input data and encode them into the hidden state. Apply the attention mechanism to the hidden state to calculate the attention weight at each time step; ; Step S430: Concatenate the hidden state output by the long short-term memory layer and the attention context vector output by the attention mechanism, and obtain the predicted future load data through the output layer ; The specific method for constructing a reinforcement learning model to optimize the unit output adjustment value of a gas-steam combined cycle power plant based on the current unit output, future load data, and adaptive time window mechanism is as follows: Step S510: Define the state space of the reinforcement learning model, and encode the predicted future load data, current unit output, and unit operation constraint conditions into the state vector of the reinforcement learning model ; The unit operation constraint conditions include the upper and lower limits of unit output, the rate of change of unit output, unit start-stop constraints, and the minimum operation time and shutdown time constraints of the unit; Step S520: Define the action space of the reinforcement learning model, and encode the output adjustment value of each unit into the action vector of the reinforcement learning model , and the action space satisfies the unit operation constraint conditions; Step S530: Define the reward function of the reinforcement learning model, and the reward function includes the unit operation cost, load tracking performance, and unit operation constraint condition indicator function; Step S540: Use a deep reinforcement learning algorithm to train the reinforcement learning model.

[0007] Step S550: Integrate the adaptive time window mechanism into the reinforcement learning model to dynamically adjust the optimization period T of the reinforcement learning model; Integrating the adaptive time window mechanism into the reinforcement learning model to dynamically adjust the optimization period T of the reinforcement learning model means that when the load fluctuation characteristic is greater than the load fluctuation threshold or the prediction error feedback is greater than the prediction error fluctuation threshold , calculate the period difference between the optimization period and the optimization adjustment step , compare the size of the period difference with the minimum value of the optimization period , and update the next optimization period to the larger value between the two; when the load fluctuation characteristic is less than the load fluctuation threshold and the prediction error feedback is less than the prediction error fluctuation threshold, calculate the period sum between the optimization period and the optimization adjustment step , compare the size of the period sum with the maximum value of the optimization period , and update the next optimization period to the smaller value between the two; otherwise, the next optimization period remains unchanged; Step S560: Use the updated decision time domain to construct the state space, action space, and reward function of the reinforcement learning model at the next time step, make a decision at the (t + 1)-th time step, and obtain the optimal unit output adjustment value.

[0008] One aspect of the present application provides an artificial intelligence-based gas-steam combined cycle unit power plant load prediction and optimal scheduling system, including: A power plant data acquisition module for real-time recording of the load data and unit output of the gas-steam combined cycle unit power plant; A time window adjustment module for extracting load fluctuation characteristics based on the load data, obtaining prediction error feedback, introducing an adaptive time window mechanism, and adjusting the time window length of the next prediction according to the current load fluctuation characteristics and prediction error feedback; A multi-time scale fusion stage for extracting multi-time scale fusion characteristics of the load data at each time step within the time window according to the updated time window length; A future load prediction module for constructing a load prediction model and predicting future load data based on the multi-time scale fusion characteristics at each time step; A unit output adjustment module for constructing a reinforcement learning model to optimize the unit output adjustment value of the gas-steam combined cycle unit power plant based on the current unit output, future load data, and adaptive time window mechanism.

[0009] One aspect of the present application provides a readable storage medium storing a computer program suitable for being loaded by a processor to execute the steps in the artificial intelligence-based gas-steam combined cycle unit power plant load prediction and optimal scheduling method.

[0010] The artificial intelligence-based gas-steam combined cycle unit power plant load prediction and optimal scheduling method and system proposed in the present application have the following advantages compared with the prior art: The present application introduces an adaptive time window mechanism, which dynamically adjusts the time window length according to the current load fluctuation characteristics and prediction error feedback, enabling the selection of the time window to adapt to the dynamic characteristics of load changes. When the load fluctuates violently or the prediction error is large, the time window is shortened to respond quickly; when the load is stable or the prediction is accurate, the time window is extended to capture long-term trends. This adaptive mechanism improves the flexibility and adaptability of load feature extraction and prediction.

[0011] The present application extracts multi-time scale fusion characteristics, comprehensively considering the characteristics of load data in the time domain, frequency domain, and time-frequency domain at different time window levels. This multi-time scale fusion method can comprehensively describe the dynamic change law of the load, including short-term fluctuations and long-term trends, and improves the accuracy and robustness of load prediction.

[0012] The present application introduces an attention mechanism into the load forecasting model, and adaptively assigns importance weights to different historical time steps according to the multi-time scale fusion features of the current time step. This attention mechanism enables the model to automatically focus on the historical information that is most helpful for the current prediction according to the dynamic characteristics of the current load, improving the pertinence and accuracy of the prediction.

[0013] The present application integrates an adaptive time window mechanism into the reinforcement learning model to dynamically adjust the optimization period, enabling the scheduling decision to be synchronized with the load prediction and adapting to the dynamic characteristics of load changes.

[0014] The present application comprehensively considers the unit operating cost, load tracking performance, and unit operating constraints in the reward function to guide the reinforcement learning model to learn a scheduling strategy that takes into account economy, reliability, and feasibility. This multi-objective optimization method can improve the operating efficiency and economic benefits while ensuring the safe and stable operation of the power plant. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 is a method flow chart of the power plant load forecasting and optimal scheduling method for a gas-steam combined cycle unit based on artificial intelligence provided by the present application; Figure 2 is an implementation method flow chart of the adaptive time window mechanism provided by the present application; Figure 3 is a method flow chart for obtaining multi-time scale fusion features provided by the present application; Figure 4 is a functional module diagram of the power plant load forecasting and optimal scheduling system for a gas-steam combined cycle unit based on artificial intelligence provided by the present application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0016] To better understand the present application, various aspects of the present application will be described in more detail with reference to the accompanying drawings. It should be understood that these detailed descriptions are only descriptions of the exemplary embodiments of the present application and do not limit the scope of the present application in any way. Throughout the specification, the same reference numerals refer to the same elements. The expression "and / or" includes any and all combinations of one or more of the associated listed items.

[0017] In the accompanying drawings, for ease of illustration, the sizes, dimensions and shapes of the elements have been slightly adjusted. The drawings are provided by way of example and are not drawn to an exact scale. As used herein, terms such as "substantially", "about" and similar terms are used as terms of approximation and not as terms of degree, and are intended to account for the inherent deviations in measured or calculated values that would be recognized by a person of ordinary skill in the art. Additionally, in this application, the order in which the steps of the processes are described does not necessarily represent the order in which these processes occur in actual operation, unless otherwise specifically limited or derivable from the context.

[0018] It should also be understood that expressions such as "comprising", "including", "having", "containing" and / or "including having" in this specification are open-ended rather than closed-ended expressions, which mean that the stated features, elements and / or components exist, but do not exclude the existence of one or more other features, elements, components and / or combinations thereof. In addition, when an expression such as "at least one of..." appears after a list of listed features, it modifies the entire list of features rather than just individual elements in the list. In addition, when describing the embodiments of this application, the use of "may" means "one or more embodiments of this application". And the term "exemplary" is intended to refer to an example or illustration.

[0019] Unless otherwise defined, all terms used herein (including engineering terms and scientific and technical terms) have the same meaning as commonly understood by a person of ordinary skill in the art to which this application pertains. It should also be understood that, unless clearly stated in this application, words defined in a commonly used dictionary should be interpreted as having a meaning consistent with their meaning in the context of the relevant art and should not be interpreted in an idealized or overly formal sense.

[0020] It should be noted that, without conflict, the embodiments in this application and the features in the embodiments may be combined with each other. The following will describe this application in detail with reference to the accompanying drawings and in combination with the embodiments.

[0021] Embodiment 1 As Figure 1 shown, the present application provides a method for power load prediction and optimal scheduling of a gas-steam combined cycle unit power plant based on artificial intelligence, including: Step S100: Record the load data and unit output of the gas-steam combined cycle unit power plant in real time; The specific method for recording the load data and unit output of the gas-steam combined cycle unit power plant in real time is: Record the load data and unit output of the gas-steam combined cycle unit power plant in real time and store them in a database, and obtain the load data for the most recent TN time steps from the database of the power plant , perform data cleaning on the load data to remove abnormal and invalid data, and interpolate missing data; among them, represents the load data at the t-th time step, and TN is the total number of time steps; Step S200: Extract load fluctuation features based on the load data, obtain prediction error feedback, introduce an adaptive time window mechanism, and adjust the time window length for the next prediction according to the current load fluctuation features and prediction error feedback; The load fluctuation features are used to measure the fluctuation degree of the load data within a time window, reflecting the change rate and amplitude of the load; The prediction error feedback is used to measure the deviation degree between the load prediction value and the actual value within a time window, reflecting the performance and applicability of the load prediction model; The specific method of extracting load fluctuation features based on the load data, obtaining prediction error feedback, introducing an adaptive time window mechanism, and adjusting the time window length for the next prediction according to the current load fluctuation features and prediction error feedback is as follows: Step S210: Divide NI types of time window levels according to different time spans. Assume that the time span of the ni-th time window level is , and calculate the number of time windows at this time window level; Preferably, the time window levels include short-term, medium-term, and long-term; The calculation formula for the number of time windows is: , where represents rounding down, and TN represents the time step of the historical load data for prediction; Step S220: For the ni-th time window level, obtain the load data of the nj-th time window therein ; The load data is expressed as , where represents the load data at the t-th time step, is the start time step of the nj-th time window at the ni-th time window level, is the length of the nj-th time window at the ni-th time window level; The time window length refers to the number of time steps included in a single time window within each time window level; Step S230: For the nj-th time window in the ni-th time window level, calculate the load fluctuation features of this time window according to the time window length and the load data of adjacent time steps in the corresponding time window ; The calculation formula for the load fluctuation features is: , where is the load data at the (t + 1)-th time step in the time window; Step S240: According to the predicted value of the load data and the true value , calculate the prediction error feedback ; The calculation formula for the prediction error feedback is: , where is the predicted value of the load data at the t-th time step; Step S250: Preset a load fluctuation threshold and a prediction error fluctuation threshold , compare the load fluctuation characteristics with the load fluctuation threshold, and the prediction error feedback with the prediction error fluctuation threshold, and update the length of the time window for the next prediction based on the adaptive time window mechanism; As Figure 2 shown, the specific method for updating the length of the time window for the next prediction based on the adaptive time window mechanism is: When the load fluctuation characteristics are greater than the load fluctuation threshold or the prediction error feedback is greater than the prediction error fluctuation threshold , calculate the time difference between the length of the time window and the adjustment step size of the time window length, compare the size of the time difference with the minimum value of the time window length, and update the length of the time window for the next prediction to the larger value between the two; when the load fluctuation characteristics are less than the load fluctuation threshold and the prediction error feedback is less than the prediction error fluctuation threshold, calculate the time sum between the length of the time window and the adjustment step size of the time window length, compare the size of the time sum with the maximum value of the time window length, and update the length of the time window for the next prediction to the smaller value between the two; otherwise, the length of the time window for the next prediction remains unchanged; The calculation formula for the length of the time window for the next prediction is: , where is the length of the time window for the previous prediction; The minimum and maximum values of the time window length are set by those skilled in the art according to the actual operation requirements and limitations of the power plant.

[0022] The load fluctuation threshold and the prediction error fluctuation threshold are set by those skilled in the art according to experience.

[0023] Step S300: Extract the multi-time-scale fusion features of the load data at each time step within the time window according to the updated time window length; As Figure 3 shown, the specific method for extracting the multi-time-scale fusion features of the load data at each time step within the time window according to the updated time window length is as follows: Step S310: Based on the length of the updated time window, for each time step, extract the load features corresponding to each time window level and time window to which it belongs. The load features include the load features of the load data in the time domain, frequency domain, and time-frequency domain; The load features of the load data in the time domain include the mean, variance, and peak-valley difference; The load features of the load data in the frequency domain are obtained by Fourier transform to get the spectrum, and the frequency domain features of the spectrum are extracted as the load features of the load data in the frequency domain; the frequency domain features include frequency distribution entropy and dominant frequency; The load features of the load data in the time-frequency domain are obtained by wavelet transform of the load data, and the time-frequency domain features are extracted as the load features of the load data in the time-frequency domain; the time-frequency domain features include energy distribution and energy centroid; Step S320: For each time step, calculate the fusion weight of the load features extracted from the nj-th time window of the ni-th time window level at this time step ; The calculation formula for the fusion weight is: , where is the control parameter; The control parameter is used to adjust the influence of the prediction error feedback on the weight; The number of time windows is the number of time windows calculated when the time span is equal to the time window step size of the current time window level.

[0024] The load data of historical time is divided into several time windows according to the predicted time window length. For each time step of each time window, the corresponding load features are extracted and feature fusion is performed. The fusion weight is the weight of each time window; Step S330: Based on the fusion weight, fuse the load features extracted from different time windows at time step t to obtain the fused load features ; The calculation formula for the fused load features is: , where is the load feature; The fused load features comprehensively integrate the feature information of the load data at different time scales.

[0025] Step S340: Integrate the fusion load characteristics at different time window levels to obtain the multi-time scale fusion characteristics at the t-th time step ; The purpose of this step is to extract the characteristics of the load from the perspectives of adaptive time scales and different domains, comprehensively characterize the change law of the load, and provide rich information for subsequent load forecasting and optimal scheduling.

[0026] Step S400: Construct a load forecasting model and predict future load data based on the multi-time scale fusion characteristics at each time step; The specific method for constructing the load forecasting model and predicting future load data based on the multi-time scale fusion characteristics at each time step is as follows: Step S410: Construct a load forecasting model with a long short-term memory neural network model as the basic model and introduce an attention mechanism. The load forecasting model includes an input layer, a long short-term memory layer, an attention mechanism layer, an output layer, and a loss function. The loss function includes the mean square error and a regularization term, and the load forecasting model is optimized and trained by minimizing the loss function; The calculation formula of the loss function is: , where represents all the learnable parameters of the load forecasting model, is the L2 regularization coefficient, represents the predicted value of the load data; All the learnable parameters of the load forecasting model include multiple learnable parameters in the input layer, the long short-term memory layer, the attention mechanism layer, and the output layer; The value of the L2 regularization coefficient is set by those skilled in the art according to experience.

[0027] Step S420: Take the multi-time scale fusion characteristics as input data and input them into the input layer. Use the long short-term memory layer to learn the temporal characteristics of the input data and encode them into the hidden state. Apply the attention mechanism to the hidden state to calculate the attention weights at each time step ; The long short-term memory layer is responsible for extracting the temporal features and long-term dependencies of the input data. Its structure includes: a forget gate, an input gate, an output gate, a candidate memory cell, a memory cell, and a hidden state. The forget gate is used to control which information in the memory cell of the previous time step should be forgotten. The input gate is used to control which information in the input data of the current time step should be added to the memory cell. The output gate is used to control which information in the memory cell state will ultimately be output to the hidden state. The candidate memory cell is used to represent the new memory information brought by the input data of the current time step. The memory cell is used to integrate the memory of the previous time step and the input of the current time step to form a new memory. The hidden state is used to represent the output of the long short-term memory layer at the current time step.

[0028] The role of the attention mechanism is to adaptively assign weights to the hidden states of different time steps according to the multi-time scale fusion features of the current time step, generating an attention context vector, which represents the aggregation of information from previous time steps.

[0029] The calculation process of the attention mechanism is as follows: , where is the attention score of the t-th time step for the t'-th time step, is the learnable attention vector, , are weight matrices, is the hidden state of the t'-th time step, is the multi-time scale fusion feature input at the t-th time step, is the attention weight of the hidden state of the t'-th time step for the prediction at the t-th time step, is the attention context vector of the t-th time step; The attention score is used to measure the importance of the hidden state of time step t' for the prediction at time step t. t represents the current time step, and t' represents the time step before the current time step; The weight matrix is used to map the hidden state of the t'-th time step to the same dimension as the attention vector. The weight matrix is used to map the multi-time scale fusion feature of the current time step to the same dimension as the attention vector.

[0030] Step S430: Concatenate the hidden state output by the long short-term memory layer and the attention context vector output by the attention mechanism, and obtain the predicted future load data through the output layer ; The calculation formula of the output layer is: , where , are learnable output layer parameters, is the hidden state of the final output of the long short-term memory layer; Step S500: Based on the current unit output, future load data, and adaptive time window mechanism, construct a reinforcement learning model to optimize the unit output adjustment value of a gas-steam combined cycle power plant; The specific method for constructing a reinforcement learning model to optimize the unit output adjustment value of a gas-steam combined cycle power plant based on the current unit output, future load data, and adaptive time window mechanism is as follows: Step S510: Define the state space of the reinforcement learning model, and encode the predicted future load data, current unit output, and unit operation constraint conditions into the state vector of the reinforcement learning model ; The unit operation constraint conditions include unit output upper and lower limit constraints, unit output change rate constraints, unit start-stop constraints, and unit minimum operation time and shutdown time constraints; The unit output upper and lower limit constraints are: , where represents the unit output of the th unit at the t-th time step, and respectively represent the minimum output limit and maximum output limit of the th unit; The unit output change rate constraints are: , where and respectively represent the maximum descent rate and maximum ascent rate of the th unit; The unit start-stop constraints are: , where is a binary variable representing the start-stop state of the th unit at the t-th time step, 1 represents running, 0 represents shutdown, represents the maximum number of start-stops allowed for the th unit within the optimization period, and T is the optimization period; The unit minimum operation time and shutdown time constraints are: , where and respectively represent the minimum operation time and minimum shutdown time of the th unit, represents the time step when the unit state changes.

[0031] The values of the minimum output limit and maximum output limit, maximum descent rate and maximum ascent rate, maximum number of start-stops, minimum operation time, and minimum shutdown time are determined by those skilled in the art according to the technical parameters of the unit.

[0032] Step S520: Define the action space of the reinforcement learning model, and encode the output adjustment value of each unit as an action vector of the reinforcement learning model , where the action space satisfies the unit operation constraint conditions; Step S530: Define the reward function of the reinforcement learning model, where the reward function includes the unit operation cost, the load tracking performance, and the unit operation constraint condition indicator function; The calculation formula of the reward function is: , where , , are weight coefficients, N represents the number of units, represents the -th unit's operation cost, represents the -th unit's output at the -th time step, represents the predicted value of the future load data at the -th time step, is the unit operation constraint condition indicator function; The predicted value of the load data at the -th time step is the load demand of the unit at this time step; The unit operation constraint condition indicator function is used to judge whether the unit output meets the unit operation constraint conditions. If , the unit operation constraint condition indicator function is equal to 1, otherwise it is equal to 0; Step S540: Use the deep reinforcement learning algorithm to train the reinforcement learning model.

[0033] The deep reinforcement learning algorithm uses the DQN algorithm; The process of training the reinforcement learning model is as follows: Use the predicted future load data, combined with the unit output, to generate the training data of the reinforcement learning model; During the training process, use the experience replay mechanism to randomly extract previous samples, update the parameters of the reinforcement learning model, improve the sample utilization efficiency and training stability; Use the ε-greedy strategy for exploration, randomly select actions with probability ε at the beginning of training, and gradually reduce the probability of random exploration as the training progresses to explore better scheduling strategies.

[0034] The rate of gradual decrease of the probability ε is set by those skilled in the art according to experience.

[0035] Step S550: Integrate the adaptive time window mechanism into the reinforcement learning model to dynamically adjust the optimization period T of the reinforcement learning model; Integrating the adaptive time window mechanism into the reinforcement learning model and dynamically adjusting the optimization period T of the reinforcement learning model means that when the load fluctuation feature is greater than the load fluctuation threshold or the prediction error feedback is greater than the prediction error fluctuation threshold calculate the period difference between the optimization period and the optimization adjustment step size compare the size of the period difference with the minimum value of the optimization period and update the next optimization period to the larger value between the two; when the load fluctuation feature is less than the load fluctuation threshold and the prediction error feedback is less than the prediction error fluctuation threshold, calculate the period sum between the optimization period and the optimization adjustment step size compare the size of the period sum with the maximum value of the optimization period and update the next optimization period to the smaller value between the two; otherwise, the next optimization period remains unchanged; Step S560: Use the updated decision time domain to construct the state space, action space, and reward function of the reinforcement learning model at the next time step, make a decision at the (t + 1)-th time step, and obtain the optimal unit output adjustment value; Based on the above steps, the adaptive time window mechanism is integrated into the optimization period adjustment of the reinforcement learning model. When the load fluctuation is large or the prediction error is large, shorten the optimization period so that the model can respond to the load change in a timely manner; when the load fluctuation is small and the prediction error is small, extend the optimization period so that the model can formulate a long-term optimization scheduling strategy. This method of dynamically adjusting the optimization period can improve the adaptability and robustness of the reinforcement learning model, making it better adapt to the load change characteristics of the gas-steam combined cycle unit power plant.

[0036] Embodiment 2 As Figure 4 shown, the gas-steam combined cycle unit power plant load prediction and optimization scheduling system based on artificial intelligence provided by the present application includes: A power plant data acquisition module for recording the load data and unit output of the gas-steam combined cycle unit power plant in real time; A time window adjustment module for extracting the load fluctuation feature based on the load data, obtaining the prediction error feedback, introducing the adaptive time window mechanism, and adjusting the time window length of the next prediction according to the current load fluctuation feature and prediction error feedback; A multi-time scale fusion stage for extracting the multi-time scale fusion features of the load data at each time step within the time window according to the updated time window length; A future load prediction module for constructing a load prediction model and predicting future load data based on the multi-time scale fusion features at each time step; The unit output adjustment module is used to construct a reinforcement learning model to optimize the unit output adjustment value of a gas-steam combined cycle unit power plant based on the current unit output, future load data, and an adaptive time window mechanism.

[0037] Embodiment 3 According to an embodiment of the present application, a readable storage medium is also provided. Computer-readable instructions are stored on the computer-readable storage medium. When the computer-readable instructions are run by a processor, the load prediction and optimal scheduling method for a gas-steam combined cycle unit power plant based on artificial intelligence described with reference to the above drawings according to the embodiments of the present application can be executed. The storage medium includes, but is not limited to, for example, volatile memory and / or non-volatile memory. Volatile memory may include, for example, random access memory (RAM) and cache memory, etc. Non-volatile memory may include, for example, read-only memory (ROM), hard disk, flash memory, etc.

[0038] In addition, according to an embodiment of the present application, the process described above with reference to the flowchart can be implemented as a computer software program. For example, the present application provides a non-transitory machine-readable storage medium storing machine-readable instructions that can be run by a processor to execute instructions corresponding to the method steps provided by the present application, such as: real-time recording of the load data and unit output of a gas-steam combined cycle unit power plant; extracting load fluctuation characteristics based on the load data, obtaining prediction error feedback, introducing an adaptive time window mechanism, and adjusting the time window length for the next prediction according to the current load fluctuation characteristics and prediction error feedback; extracting multi-time scale fusion characteristics of the load data at each time step within the time window according to the updated time window length; constructing a load prediction model and predicting future load data based on the multi-time scale fusion characteristics at each time step; constructing a reinforcement learning model based on the current unit output, future load data, and an adaptive time window mechanism to optimize the unit output adjustment value of a gas-steam combined cycle unit power plant. When the computer program is executed by a central processing unit (CPU), the above functions defined in the method of the present application are executed.

[0039] The method, apparatus, and device of the present application can be implemented in many ways. For example, the method, apparatus, and device of the present application can be implemented by software, hardware, firmware, or any combination of software, hardware, and firmware. The above order of steps for the method is only for illustration, and the steps of the method of the present application are not limited to the above specific described order unless otherwise specifically stated. In addition, in some embodiments, the present application can also be implemented as a program recorded in a recording medium, and these programs include machine-readable instructions for implementing the method according to the present application. Therefore, the present application also covers a recording medium storing a program for executing the method according to the present application.

[0040] In addition, parts of the above technical solutions provided in the embodiments of the present application that have the same implementation principle as the corresponding technical solutions in the prior art are not described in detail to avoid excessive elaboration.

[0041] As described above in the specific embodiments, the purpose, technical solutions, and beneficial effects of the present invention have been further described in detail. It should be understood that the above are only specific embodiments of the present invention and are not used to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A method for load forecasting and optimal dispatching of a gas-steam combined cycle power plant based on artificial intelligence, characterized in that: include: Real-time recording of load data and unit output of gas-steam combined cycle power plants; Extract load fluctuation characteristics based on load data, obtain forecast error feedback, introduce an adaptive time window mechanism, and adjust the time window length of the next forecast according to the current load fluctuation characteristics and forecast error feedback; Extract multi-time scale fusion features of load data at each time step in the time window according to the updated time window length; Construct a load forecasting model to predict future load data based on the multi-time scale fusion features of each time step; Based on the current unit output, future load data and adaptive time window mechanism, a reinforcement learning model is constructed to optimize the unit output adjustment value of the gas-steam combined cycle power plant.

2. The method for load forecasting and optimal scheduling of a gas-steam combined cycle power plant based on artificial intelligence according to claim 1, characterized in that: The specific method of extracting load fluctuation characteristics based on load data, obtaining prediction error feedback, introducing an adaptive time window mechanism, and adjusting the time window length of the next prediction according to the current load fluctuation characteristics and prediction error feedback is as follows: Divide NI time window levels according to different time spans. Assume that the time span of the ni-th time window level is , calculate the number of time windows at this time window level; For the ni-th time window level, obtain the load data of the nj-th time window ; For the njth time window in the nith time window level, the load fluctuation characteristics of the time window are calculated according to the length of the time window and the load data of the adjacent time steps in the corresponding time window. ; Predicted value based on load data With the true value , calculate the prediction error feedback ; Preset load fluctuation threshold and prediction error fluctuation threshold , compare the load fluctuation characteristics with the load fluctuation threshold, the prediction error feedback with the prediction error fluctuation threshold, and update the length of the time window for the next prediction based on the adaptive time window mechanism.

3. The method for load forecasting and optimal scheduling of a gas-steam combined cycle power plant based on artificial intelligence according to claim 2, characterized in that: The specific method for updating the length of the time window for the next prediction based on the adaptive time window mechanism is: When the load fluctuation characteristic is greater than the load fluctuation threshold When the prediction error feedback is greater than the prediction error fluctuation threshold When calculating the length of the time window and the adjustment step of the time window length The time difference between the two is compared with the minimum value of the time window length. The size between them is updated to the larger value between the two; when the load fluctuation characteristic is less than the load fluctuation threshold and the prediction error feedback is less than the prediction error fluctuation threshold, the length of the time window is calculated and the adjustment step of the time window length is The time between the two, compare the time and the maximum value of the time window length If the value is smaller than the value between the two, the length of the next predicted time window is updated to the smaller value between the two; otherwise, the length of the next predicted time window remains unchanged.

4. The method for load forecasting and optimal scheduling of a gas-steam combined cycle power plant based on artificial intelligence according to claim 3, characterized in that: The specific method for extracting the multi-time scale fusion features of the load data of each time step in the time window according to the updated time window length is: Based on the length of the updated time window, for each time step, extract the load characteristics corresponding to each time window level and the time window to which it belongs, wherein the load characteristics include the load characteristics of the load data in the time domain, the frequency domain and the time-frequency domain; For each time step, calculate the fusion weight of the load feature extracted by the nj-th time window at the ni-th time window level at that time step ; Based on the fusion weight, the load features extracted from different time windows at time step t are fused to obtain the fused load features ; Combining the fusion load characteristics of different time window levels, the multi-time scale fusion characteristics of the t-th time step are obtained. .

5. The method for load forecasting and optimal scheduling of a gas-steam combined cycle power plant based on artificial intelligence according to claim 4, characterized in that: The specific method of constructing the load forecasting model and predicting future load data based on the multi-time scale fusion features of each time step is: Construct a load forecasting model, using a long short-term memory neural network model as a basic model and introducing an attention mechanism. The load forecasting model includes an input layer, a long short-term memory layer, an attention mechanism layer, an output layer and a loss function. The loss function includes a mean square error and a regularization term. The load forecasting model is optimized and trained by minimizing the loss function. The multi-time scale fusion features are input as input data to the input layer, the long short-term memory layer is used to learn the temporal features of the input data and encode them into the hidden state, the attention mechanism is applied to the hidden state, and the attention weight of each time step is calculated. ; The hidden state output by the long short-term memory layer and the attention context vector output by the attention mechanism are concatenated to obtain the predicted future load data through the output layer. .

6. The method for load forecasting and optimal scheduling of a gas-steam combined cycle power plant based on artificial intelligence according to claim 5, characterized in that: The specific method of constructing a reinforcement learning model to optimize the unit output adjustment value of a gas-steam combined cycle unit power plant based on the current unit output, future load data and an adaptive time window mechanism is as follows: Step S510: define the state space of the reinforcement learning model, and encode the predicted future load data, current unit output, and unit operation constraints into the state vector of the reinforcement learning model ; Step S520: define the action space of the reinforcement learning model, and encode the output adjustment value of each unit into the action vector of the reinforcement learning model , the action space satisfies the unit operation constraints; Step S530: defining a reward function of the reinforcement learning model, wherein the reward function includes a unit operation cost, a load tracking performance, and a unit operation constraint condition indication function; Step S540: Use a deep reinforcement learning algorithm to train a reinforcement learning model. Step S550: Integrate the adaptive time window mechanism into the reinforcement learning model, and dynamically adjust the optimization period T of the reinforcement learning model; Step S560: Use the updated decision time domain to construct the state space, action space and reward function of the reinforcement learning model in the next time step, make a decision in the t+1th time step, and obtain the optimal unit output adjustment value.

7. The method for load forecasting and optimal scheduling of a gas-steam combined cycle power plant based on artificial intelligence according to claim 6, characterized in that: The unit operation constraints include unit output upper and lower limit constraints, unit output change rate constraints, unit start and stop constraints, and unit minimum operating time and shutdown time constraints.

8. The method for load forecasting and optimal scheduling of a gas-steam combined cycle power plant based on artificial intelligence according to claim 7, characterized in that: The adaptive time window mechanism is integrated into the reinforcement learning model, and the optimization period T of the reinforcement learning model is dynamically adjusted: when the load fluctuation characteristic is greater than the load fluctuation threshold When the prediction error feedback is greater than the prediction error fluctuation threshold When calculating the optimization cycle and optimization adjustment step The cycle difference between the two is compared with the minimum value of the optimized cycle. The size between them is updated to the larger value between the two; when the load fluctuation characteristic is less than the load fluctuation threshold and the prediction error feedback is less than the prediction error fluctuation threshold, the optimization cycle and optimization adjustment step are calculated. The sum of the periods between and comparing the sum of the periods with the maximum value of the optimization period If the value between them is large, the next optimization cycle is updated to the smaller value between them; otherwise, the next optimization cycle remains unchanged.

9. A gas-steam combined cycle unit power plant load forecasting and optimization scheduling system based on artificial intelligence, which is implemented based on the gas-steam combined cycle unit power plant load forecasting and optimization scheduling method based on artificial intelligence as described in any one of claims 1 to 8, characterized in that: include: Power plant data acquisition module, used to record the load data and unit output of the gas-steam combined cycle power plant in real time; The time window adjustment module is used to extract load fluctuation characteristics based on load data, obtain prediction error feedback, introduce an adaptive time window mechanism, and adjust the time window length of the next prediction according to the current load fluctuation characteristics and prediction error feedback; The multi-time scale fusion stage is used to extract the multi-time scale fusion features of the load data at each time step in the time window according to the updated time window length; The future load forecasting module is used to build a load forecasting model and predict future load data based on the multi-time scale fusion features of each time step; The unit output adjustment module is used to build a reinforcement learning model to optimize the unit output adjustment value of the gas-steam combined cycle power plant based on the current unit output, future load data and adaptive time window mechanism.

10. A readable storage medium, characterized in that: The readable storage medium stores a computer program, which is suitable for being loaded by a processor to execute the steps in the artificial intelligence-based gas-steam combined cycle unit power plant load forecasting and optimization scheduling method as described in any one of claims 1-8.

Images