Thermal power unit peak load intelligent cruise control method based on big data analysis

Abstract

The present application relates to the technical field of deep peak regulation of thermal power generating units, and more particularly to a thermal power generating unit peak regulation load intelligent cruise control method based on big data analysis, which is used to solve the problem that the prior art cannot adaptively adjust the output of the thermal power generating unit based on real-time monitoring, cannot select the optimal action according to real-time data, and reduces the stability, economy and operation efficiency of the power grid; the present application uses a reinforcement learning algorithm to optimize the peak regulation control strategy, adaptively adjusts the output of the thermal power generating unit based on real-time monitoring, balances system safety and economy by comprehensively considering the frequency stability, output constraint, economy and unit state to design a reward function, the experience replay and target network mechanism of the DQN algorithm improve the learning efficiency, realize fast convergence and continuous optimization, and after training, the system selects the optimal action according to real-time data, which significantly improves the stability, economy and operation efficiency of the power grid.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of deep peak regulation of thermal power generating units, in particular to a method for intelligent cruise control of peak load of thermal power generating units based on big data analysis. BACKGROUND

[0002] With the continuous growth of power demand and the rapid development of new energy, the role of thermal power generating units in peak regulation of power grids is becoming increasingly important. Traditional peak regulation control methods for thermal power generating units often rely on manual experience and fixed parameter settings, which are difficult to adapt to complex and variable power grid load demands and unit operating conditions.

[0003] The patent application with publication number CN117498455A discloses an intelligent deep peak regulation coordination method and system. The operating information of a thermal power generating unit is obtained, and it is determined whether the thermal power generating unit meets the peak regulation conditions according to the operating information. If so, a peak regulation identifier is generated for the thermal power generating unit. The unit load of the thermal power generating unit is collected based on the peak regulation identifier. The peak regulation identifier is analyzed to determine the peak regulation state of the thermal power generating unit. When the peak regulation state of the thermal power generating unit is deep peak regulation, the coal consumption cost and the loss cost of the thermal power generating unit in the deep peak regulation state are calculated. The unit load of the thermal power generating unit is adjusted according to the coal consumption cost and the loss cost of the thermal power generating unit to obtain a deep peak regulation scheme for the thermal power generating unit. The present application adjusts the unit load according to the coal consumption cost and the loss cost of the thermal power generating unit, which can realize intelligent peak regulation of the thermal power generating unit, improve the control precision of deep peak regulation and the economy of the thermal power generating unit, and reduce the energy loss of the thermal power generating unit.

[0004] However, the above-mentioned reference patent adjusts the unit load intelligently by analyzing the coal consumption and loss cost of the thermal power generating unit to realize precise and economically efficient deep peak regulation control, but cannot realize high-precision power grid load prediction, cannot effectively support peak regulation strategies, cannot realize precise allocation and dynamic adjustment of the output of the thermal power generating unit, cannot optimize the output strategy in real time according to actual load changes, reduces the response speed, control precision and safety of the system, and cannot adaptively adjust the output of the thermal power generating unit based on real-time monitoring, cannot select the optimal action according to real-time data, and reduces the stability, economy and operation efficiency of the power grid.

[0005] Therefore, we propose a method for intelligent cruise control of peak load of thermal power generating units based on big data analysis to solve the above-mentioned problems. SUMMARY

[0006] The purpose of the present application is to provide a thermal power unit peak load intelligent cruise control method based on big data analysis, which solves the problem that the prior art cannot realize high-precision power grid load prediction, cannot effectively support peak shaving strategy, cannot realize precise allocation and dynamic adjustment of thermal power unit output, cannot optimize output strategy in real time according to actual load changes, reduces the response speed, control accuracy and safety of the system, and cannot adaptively adjust the thermal power unit output based on real-time monitoring, cannot select the optimal action according to real-time data, and reduces the stability, economy and operation efficiency of the power grid.

[0007] The purpose of the present application is realized by the following technical solutions:

[0008] The thermal power unit peak load intelligent cruise control method based on big data analysis comprises the following steps:

[0009] Step one: Real-time collection of operation data of the thermal power unit, and pre-processing operation of the collected operation data;

[0010] Step two: Collection of historical operation data of the thermal power unit, construction of a power grid load prediction model, and prediction of the power grid load in a future period of time through the model;

[0011] Step three: Real-time monitoring of key operation parameters of the thermal power unit, and monitoring and evaluation of the operation state of the thermal power unit;

[0012] Step four: Formulation of a corresponding peak shaving control strategy according to the power grid load prediction result and the thermal power unit state monitoring result;

[0013] Step five: Optimization of the peak shaving control strategy by using a reinforcement learning algorithm based on the thermal power unit state monitoring result.

[0014] As a preferred embodiment of the present application, the specific process of the step two of predicting the power grid load in a future period of time through the model is as follows:

[0015] The historical operation data of the thermal power unit are obtained, the operation data including the boiler pressure, the boiler temperature, the feedwater flow, the turbine speed, the generator voltage, the generator current, the fuel flow, the power grid frequency, the power grid voltage, the load demand, the environmental temperature and the environmental humidity, a collection period is generated, and the collection period is equally divided into a plurality of collection time periods;

[0016] The load demand change rate of the thermal power unit in the plurality of collection time periods is obtained, the load demand change rate representing the ratio between the load demand change amount and the corresponding time period length, a set A of load demand change rates is constructed in this way, and the average value of the difference between the maximum subset and the minimum subset in the set A is recorded as a load demand change rate difference FBC;

[0017] The method for calculating the load demand change rate difference FBC can also be used to obtain an ambient temperature change rate difference WBC, a power grid frequency change rate difference PBC, and a power grid voltage change rate difference DBC.

[0018] The load demands of the thermal power generating unit in multiple collection time periods are obtained, and the obtained multiple load demands are calculated to obtain an arithmetic mean value, which is denoted as an average load demand PFX.

[0019] The method for calculating the average load demand PFX can also be used to obtain an average ambient temperature PHW, an average power grid frequency PDP, and an average power grid voltage PDY.

[0020] As a preferred embodiment of the present application, the load demand change rate difference FBC, the ambient temperature change rate difference WBC, the power grid frequency change rate difference PBC, the power grid voltage change rate difference DBC, the average load demand PFX, the average ambient temperature PHW, the average power grid frequency PDP, and the average power grid voltage PDY are combined to construct a power grid load prediction matrix DFJ, the power grid load prediction matrix DFJ is used as an input of a machine learning model, and a power grid load in a future period of time corresponding to each group of power grid load prediction matrix DFJ is used as an output of the machine learning model, the power grid load in the future period of time is taken as a prediction target, the sum of prediction errors of all training data is taken as a training target, the machine learning model is trained until the sum of prediction errors converges, and a power grid load prediction model is obtained.

[0021] Real-time operation data of the thermal power generating unit are obtained, the real-time operation data are converted into a corresponding power grid load prediction matrix DFJ, the power grid load prediction matrix DFJ is input into the power grid load prediction model, and a power grid load prediction value in a future period of time is obtained through the power grid load prediction model t = 1, 2, …, T, T is a prediction step.

[0022] As a preferred embodiment of the present application, the specific process of the step three of monitoring and evaluating the operation state of the thermal power generating unit is as follows:

[0023] Key operation parameters of the thermal power generating unit are obtained, the key operation parameters include a boiler pressure, a turbine speed, a fuel flow, and a load demand, a monitoring period is generated, and the monitoring period is equally divided into multiple monitoring time periods;

[0024] A boiler pressure imbalance value of the thermal power generating unit in the monitoring period is obtained, and the boiler pressure imbalance value is denoted as GYS, the boiler pressure imbalance value represents a ratio between a portion in which a boiler pressure variation difference in each monitoring time period is greater than a preset boiler pressure variation difference threshold value and the boiler pressure variation difference, and the boiler pressure variation difference represents a difference between a maximum value and a minimum value of the boiler pressure;

[0025] Similarly, the turbine speed imbalance value QZS, fuel flow imbalance value RLS, and load demand imbalance value FXS can be obtained using the same method as calculating the boiler pressure imbalance value GYS.

[0026] In a preferred embodiment of the present invention, the boiler pressure imbalance value GYS, turbine speed imbalance value QZS, fuel flow imbalance value RLS, and load demand imbalance value FXS are obtained, and the operating status evaluation coefficient YZP is calculated using the following formula:

[0027]

[0028] Where d1, d2, d3, and d4 are all preset scaling factor coefficients, d4 > d3 > d2 > d1 > 0. The operating status evaluation coefficient YZP is compared with the preset first operating status evaluation coefficient threshold and the preset second operating status evaluation coefficient threshold. The preset first operating status evaluation coefficient threshold is less than the preset second operating status evaluation coefficient threshold.

[0029] If the operating status evaluation coefficient YZP is less than the preset first operating status evaluation coefficient threshold, it indicates that the thermal power unit is in excellent operating status.

[0030] If the operating status evaluation coefficient YZP is greater than or equal to the preset first operating status evaluation coefficient threshold, and the operating status evaluation coefficient YZP is less than the preset second operating status evaluation coefficient threshold, it indicates that the thermal power unit is operating normally.

[0031] If the operating status evaluation coefficient YZP is greater than or equal to the preset second operating status evaluation coefficient threshold, it indicates that the thermal power unit is in an abnormal operating state.

[0032] As a preferred embodiment of the present invention, the specific process of formulating the corresponding peak-shaving control strategy in step four is as follows:

[0033] The power grid load forecast results and thermal power unit condition monitoring results were obtained. The power grid load forecast results are the predicted power grid load values ​​for a future period of time. The status monitoring results of the thermal power unit indicate that the thermal power unit is in excellent, normal, or abnormal operating condition.

[0034] Assuming there are N thermal power units, obtain the maximum output P of each unit. max,i and actual grid load L st The operating status S of the thermal power unit i Let Si ∈ {1, 0.5, 0}, where 1 represents excellent operating condition of the thermal power unit, 0.5 represents normal operating condition, and 0 represents abnormal operating condition. Let S represent the operating condition of the thermal power unit. i Convert to weight W i :

[0035]

[0036] The content of the formulated peak shaving control strategy is to adjust the output distribution of each thermal power unit. The weighted average method is used to calculate the initial output distribution:

[0037] The total weighted output capacity is calculated:

[0038] The output proportion of each unit is calculated:

[0039] The initial output of each unit is calculated: P i = R i × L t .

[0040] As a preferred embodiment of the present application, the PI control algorithm is then used to adjust the output of the thermal power unit, to compensate for the prediction error and meet the safety constraints:

[0041] Prediction error calculation:

[0042] PI control algorithm: ΔP i = K p · e t + K i ∫e t dt;

[0043] Where K p and K i are proportional gain and integral gain, which need to be adjusted according to the system characteristics and load prediction accuracy;

[0044] Output adjustment: P i new = P i + ΔP i ;

[0045] Safety constraints and output limits: P i final = min(P i new , P max,i );

[0046] P i final ≤ P i + ΔP rate × Δt, where ΔP rate is the limit of the change rate of the output of the thermal power unit, which represents the maximum value of the allowed change of the unit output per unit time, and Δt is the control period;

[0047] P i finalThe data is sent to each thermal power unit for control, adjusting the output distribution of each unit. Within each control cycle, data is reacquired and W is recalculated. i Re-implement PI control and adjust K according to the actual situation. p and K i The output of thermal power units is adjusted in real time to track changes in actual load.

[0048] As a preferred embodiment of the present invention, the specific process of optimizing the peak shaving control strategy using a reinforcement learning algorithm in step five is as follows:

[0049] The specific steps for optimizing the peak-shaving control strategy using reinforcement learning algorithms are as follows: The results of the thermal power unit status monitoring are obtained, indicating whether the thermal power unit's operating status is excellent, normal, or abnormal.

[0050] First, define a state space Z, which contains the following information:

[0051] Operating status S of each thermal power unit i Current power grid load L st The current output P of each thermal power unit i Forecast values ​​of power grid load in the future period The state space Z can be represented as a vector:

[0052]

[0053] Let's define the action space A, which represents the adjustment of the output of each thermal power unit. The action is represented as a vector: a = [ΔP1, ΔP2, ..., ΔP N ], where ΔP i This indicates the amount of output adjustment for thermal power unit i.

[0054] In a preferred embodiment of the present invention, a reward function R is then defined, which comprehensively considers multiple factors:

[0055] Frequency stability: The reward function should encourage maintaining system frequency stability, using the square of the frequency deviation as a penalty term: R f =-K f (Δf) 2 Where Δf is the frequency deviation, k f These are the weighting coefficients;

[0056] Output constraints: If the output of a unit exceeds its upper limit or falls below its lower limit, a penalty will be imposed. Where P i,min and P i,max These represent the lower and upper limits of unit i's output, respectively; 1 represents the indicator function; k c These are the weighting coefficients;

[0057] Economy: using fuel consumption rate as a penalty term, the relationship model between fuel consumption rate and unit output needs to be established, assuming that the relationship is F i i , then the penalty term is:

[0058] where k e is the weight coefficient;

[0059] Unit state: preferentially use the unit with excellent running state, and give rewards:

[0060] where k s is the weight coefficient;

[0061] The total reward function can be expressed as: R=R f +R c +R e +R s .

[0062] As a preferred embodiment of the present application, the optimal peak shaving control strategy is finally learned using the DQN algorithm, and the core of the DQN algorithm is to learn a Q network which approximates the Q function Q(s,a; θ), wherein θ is the network parameter, and the training process of the DQN algorithm is as follows:

[0063] T1: randomly initialize Q network parameters θ and target Q network parameters θ'= θ, and create an experience replay pool D;

[0064] T2: repeat the following steps until convergence:

[0065] T21: obtain a sample (s, a, r, s') from the environment, wherein s is the current state, a is the action taken, r is the reward obtained, and s' is the next state;

[0066] T22: add the sample (s, a, r, s') to the experience replay pool D;

[0067] T23: randomly sample a batch of samples from the experience replay pool D;

[0068] T24: for each sample (s i , a i , r i , s′ i ), calculate the target value y i :

[0069] y i = r i + γmax a Q(s′ i ​where gamma is a discount factor;

[0070] T25: updating the Q-network parameters theta using a mean squared error loss function:

[0071] Gradient descent can be used to minimize the loss function;

[0072] T26: periodically copying the parameters of the Q-network to the target Q-network;

[0073] T3: after training, the trained Q-network can be used to select actions, given a state s, select the action a that maximizes Q(s, a; theta);

[0074] After training, the learned Q-network can be deployed to the actual system, according to the real-time state monitoring results, the best action is selected for peak shaving control.

[0075] Compared with the prior art, the advantages of the present application are:

[0076] (1) In the present application, by collecting historical operation data, a prediction matrix DFJ containing key features is constructed, and a machine learning model is trained to error convergence, realizing high-precision power grid load prediction, which improves the prediction accuracy, system adaptability, effectively supports the peak shaving strategy, and optimizes the management and operation efficiency of thermal power units;

[0077] (2) In the present application, by combining the power grid load prediction results and the thermal power unit state monitoring results, using the weighted average method and the PI control algorithm, the accurate distribution and dynamic adjustment of the thermal power unit output are realized, which not only can effectively compensate the prediction error, meet the safety constraints, but also can optimize the output strategy in real time according to the actual load change, improve the response speed, control precision and safety of the system, optimize the economy and efficiency, and ensure efficient and stable power supply;

[0078] (3) In the present application, the reinforcement learning algorithm is used to optimize the peak shaving control strategy, the thermal power unit output is adjusted adaptively based on real-time monitoring, the reward function is designed by comprehensively considering the frequency stability, output constraint, economy and unit state, the system safety and economy are balanced, the experience replay and target network mechanism of DQN algorithm improve the learning efficiency, realize fast convergence and continuous optimization, after training, the system selects the optimal action according to real-time data, significantly improves the stability, economy and operation efficiency of the power grid. BRIEF DESCRIPTION OF DRAWINGS

[0079] Figure 1 The flowchart of the intelligent cruise control method for thermal power unit peak shaving load in the present application;

[0080] Figure 2The logical flowchart in the first embodiment of the present application is shown in the figure.

[0081] Figure 3 The flowchart of the training process of the DQN algorithm in the present application is shown in the figure.

[0082] Figure 4 The sub-step flowchart of the T2 step in the training process of the DQN algorithm in the present application is shown in the figure. DETAILED DESCRIPTION

[0083] The technical solutions in the embodiments of the present application will be described clearly and completely in the following with reference to the accompanying drawings; obviously, the described embodiments are only a part of the embodiments of the present application, but not all the embodiments; based on the embodiments in the present application, all the other embodiments obtained by a person of ordinary skill in the art without any creative work fall within the protection scope of the present application.

[0084] Embodiment one: as shown in the figure, the intelligent cruise control method for thermal power unit peak load based on big data analysis comprises the following steps: Figure 1 and Figure 2 The intelligent cruise control method for thermal power unit peak load based on big data analysis comprises the following steps:

[0085] Step one: real-time collection of the operation data of the thermal power unit, and pre-processing operation on the collected operation data, the pre-processing operation including but not limited to data cleaning, filtering processing and normalization processing;

[0086] Through the data cleaning, filtering and normalization processing, the accuracy, stability and consistency of the operation data of the thermal power unit are ensured, the model performance and calculation efficiency are improved, the adaptive ability and analysis accuracy of the system are enhanced, reliable data support is provided for the load prediction and peak shaving strategy, and efficient and accurate unit management and optimization are realized.

[0087] Step two: collection of the historical operation data of the thermal power unit, construction of a power grid load prediction model, and prediction of the power grid load in a future period of time through the model;

[0088] The specific process of step two for predicting the power grid load in a future period of time through the model is as follows:

[0089] The historical operation data of the thermal power unit are obtained, the operation data including the boiler pressure, the boiler temperature, the feedwater flow, the turbine speed, the generator voltage, the generator current, the fuel flow, the power grid frequency, the power grid voltage, the load demand, the environment temperature and the environment humidity, a collection period is generated, and the collection period is equally divided into multiple collection time periods;

[0090] Obtain the load demand change rate of the thermal power unit in multiple collection periods, which represents the ratio of the load demand change amount to the corresponding time period length, to construct a set A of load demand change rates, and record the average of the difference between the maximum subset and the minimum subset in set A as the load demand change rate difference FBC.

[0091] The method for calculating the load demand change rate difference FBC can also be used to obtain the environmental temperature change rate difference WBC, the grid frequency change rate difference PBC, and the grid voltage change rate difference DBC.

[0092] Obtain the load demand of the thermal power unit in multiple collection periods, and calculate the arithmetic mean of the obtained multiple load demands, recording the arithmetic mean of the multiple load demands as the average load demand PFX.

[0093] The method for calculating the average load demand PFX can also be used to obtain the average environmental temperature PHW, the average grid frequency PDP, and the average grid voltage PDY.

[0094] Combine the load demand change rate difference FBC, the environmental temperature change rate difference WBC, the grid frequency change rate difference PBC, the grid voltage change rate difference DBC, the average load demand PFX, the average environmental temperature PHW, the average grid frequency PDP, and the average grid voltage PDY to construct a grid load prediction matrix DFJ. Use the grid load prediction matrix DFJ as the input of the machine learning model, and use the grid load in the future period corresponding to each group of grid load prediction matrix DFJ as the output of the machine learning model. Take the grid load in the future period as the prediction target, and take minimizing the sum of prediction errors of all training data as the training target. Train the machine learning model until the sum of prediction errors converges, and stop training to obtain a grid load prediction model.

[0095] Obtain the real-time operation data of the thermal power unit, convert it into the corresponding grid load prediction matrix DFJ, input it into the grid load prediction model, and obtain the grid load prediction value in the future period through the grid load prediction model t = 1, 2, …, T, T is the prediction step length.

[0096] By collecting historical operation data, a prediction matrix DFJ containing key features (such as load demand and environmental change rate difference) is constructed, and a machine learning model is trained until the error converges, achieving high-precision grid load prediction. This method improves the prediction accuracy and system adaptability, effectively supports the peak shaving strategy, and optimizes the management and operation efficiency of the thermal power unit.

[0097] Step three: Real-time monitoring of key operating parameters of the thermal power unit, monitoring and evaluating the operating state of the thermal power unit.

[0098] The specific process of step three of monitoring and evaluating the operation state of the thermal power generating unit is as follows:

[0099] Key operation parameters of the thermal power generating unit are acquired, the key operation parameters including boiler pressure, turbine speed, fuel flow and load demand, a monitoring period is generated, and the monitoring period is equally divided into multiple monitoring time periods;

[0100] The boiler pressure imbalance value of the thermal power generating unit in the monitoring period is acquired, and the boiler pressure imbalance value is marked as GYS, the boiler pressure imbalance value representing the ratio between the part of each monitoring time period in which the boiler pressure variation difference is greater than the preset boiler pressure variation difference threshold and the boiler pressure variation difference, the boiler pressure variation difference representing the difference between the maximum value and the minimum value of the boiler pressure;

[0101] The turbine speed imbalance value QZS, the fuel flow imbalance value RLS and the load demand imbalance value FXS can be obtained by using the same method as calculating the boiler pressure imbalance value GYS;

[0102] The boiler pressure imbalance value GYS, the turbine speed imbalance value QZS, the fuel flow imbalance value RLS and the load demand imbalance value FXS are acquired, and the operation state evaluation coefficient YZP is calculated by the following formula:

[0103]

[0104] Wherein d1, d2, d3 and d4 are all preset proportional factor coefficients, d4>d3>d2>d1>0, the operation state evaluation coefficient YZP is compared with the preset first operation state evaluation coefficient threshold and the preset second operation state evaluation coefficient threshold, and the preset first operation state evaluation coefficient threshold is less than the preset second operation state evaluation coefficient threshold:

[0105] If the operation state evaluation coefficient YZP is less than the preset first operation state evaluation coefficient threshold, it indicates that the operation state of the thermal power generating unit is excellent;

[0106] If the operation state evaluation coefficient YZP is greater than or equal to the preset first operation state evaluation coefficient threshold and less than the preset second operation state evaluation coefficient threshold, it indicates that the operation state of the thermal power generating unit is normal;

[0107] If the operation state evaluation coefficient YZP is greater than or equal to the preset second operation state evaluation coefficient threshold, it indicates that the operation state of the thermal power generating unit is abnormal;

[0108] By monitoring the key operation parameters of the thermal power generating unit (such as the boiler pressure, the turbine speed, etc.) in real time, calculating the imbalance value and constructing the operation state evaluation coefficient YZP, the accurate evaluation (excellent, normal or abnormal) of the operation state of the thermal power generating unit is realized. The method can not only provide early warning to prevent faults from occurring, but also can optimize the operation strategy according to the evaluation result, improve the safety, stability and economy of the system and ensure efficient operation.

[0109] Step four: according to the power grid load prediction result and the thermal power generating unit state monitoring result, a corresponding peak regulation control strategy is formulated;

[0110] The specific process of step four of formulating the corresponding peak regulation control strategy is as follows:

[0111] After the power grid load prediction result and the thermal power generating unit state monitoring result are obtained, the power grid load prediction result is the power grid load prediction value in a future period of time The thermal power generating unit state monitoring result is that the operation state of the thermal power generating unit is excellent, normal or abnormal;

[0112] Suppose there are N thermal power generating units, the maximum output P max,i and the actual power grid load L st of each unit are obtained, i the operation state S i of the thermal power generating unit is represented by a numerical value: Si∈{1, 0.5, 0}, 1 represents that the operation state of the thermal power generating unit is excellent, 0.5 represents that the operation state of the thermal power generating unit is normal, and 0 represents that the operation state of the thermal power generating unit is abnormal, the operation state S i of the thermal power generating unit is converted into a weight W i :

[0113]

[0114] The content of the formulated peak regulation control strategy is the output distribution of each thermal power generating unit. The weighted average method is used to calculate the initial output distribution first:

[0115] The total weighted output capacity is calculated:

[0116] The output proportion of each unit is calculated:

[0117] The initial output P i of each unit is calculated: t ;

[0118] Then, the PI control algorithm is used to adjust the output of the thermal power generating unit, compensate for the prediction error and meet the safety constraints:

[0119] The prediction error is calculated:

[0120] PI control algorithm: ΔP i = K p ·e t + K i ∫e t dt;

[0121] where K p and K i are proportional and integral gains, which need to be adjusted according to system characteristics and load prediction accuracy;

[0122] Output adjustment: P i new = P i + ΔP i ;

[0123] Safety constraints and output limits: P i final = min(P i new , P max，i )(not exceeding maximum output);

[0124] P i final ≤ P i + ΔP rate × Δt(satisfying output rate limit), where ΔP rate is the limit of thermal power unit output rate, representing the maximum allowed change in unit output per unit time, and Δt is the control period. Considering the operating state of the thermal power unit, if it is in a maintenance state or other abnormal state, its P i final should be set to 0 or a very small value;

[0125] Send P i final to each thermal power unit for control, adjust the output distribution of each thermal power unit, and in each control period, reacquire data, recalculate W i , re-PI control, and adjust K p and K i according to actual conditions, and adjust the thermal power unit output in real time to track actual load changes;

[0126] By combining the grid load prediction results and thermal power unit state monitoring results, using the weighted average method and PI control algorithm, precise distribution and dynamic adjustment of thermal power unit output are achieved. This method not only effectively compensates for prediction errors and meets safety constraints, but also optimizes output strategies in real time according to actual load changes, improves system response speed, control accuracy and safety, optimizes economy and efficiency, and ensures efficient and stable power supply.

[0127] Example 2: The technical solution of this embodiment of the invention differs from that of Example 1 in that:

[0128] like Figure 1 , Figure 3 and Figure 4 As shown, step five: Based on the status monitoring results of thermal power units, optimize the peak-shaving control strategy using reinforcement learning algorithms;

[0129] Step five involves optimizing the peak-shaving control strategy using reinforcement learning algorithms, and the specific process is as follows:

[0130] The specific steps for optimizing the peak-shaving control strategy using reinforcement learning algorithms are as follows: The results of the thermal power unit status monitoring are obtained, indicating whether the thermal power unit's operating status is excellent, normal, or abnormal.

[0131] First, define a state space Z, which contains the following information:

[0132] Operating status S of each thermal power unit i Current power grid load L st The current output P of each thermal power unit i Forecast values ​​of power grid load in the future period The state space Z can be represented as a vector:

[0133]

[0134] Let's define the action space A, which represents the adjustment of the output of each thermal power unit. The action is represented as a vector: a = [ΔP1, ΔP2, ..., ΔP N ], where ΔP i This represents the output adjustment amount for thermal power unit i. To limit the operating range, a maximum adjustment amount ΔP can be set. max Then -ΔP max ≤ΔP i ≤ΔP max ;

[0135] Next, we define the reward function R, which takes into account multiple factors:

[0136] Frequency stability: The reward function should encourage maintaining system frequency stability, using the square of the frequency deviation as a penalty term: R f =-K f (Δf) 2 Where Δf is the frequency deviation, k f These are the weighting coefficients;

[0137] Output constraints: If the output of a unit exceeds its upper limit or falls below its lower limit, a penalty will be imposed. Where Pi,min and P i,max are the lower and upper limits of the output of unit i, respectively, 1 is an indicator function, k c is a weight coefficient;

[0138] Economy: Using fuel consumption rate as the penalty term, the relationship model between fuel consumption rate and unit output needs to be established, assuming that the relationship is F i (P i ), then the penalty term is:

[0139] where k e is a weight coefficient;

[0140] Unit state: Preferentially use units with excellent operating state, which can be rewarded:

[0141] where k s is a weight coefficient;

[0142] The total reward function can be represented as: R = R f + R c + R e + R s ;

[0143] Finally, the optimal peak shaving control strategy is learned using the DQN algorithm, the core of the DQN algorithm is to learn a Q network, which approximates the Q function Q(s, a; θ), where θ is the network parameter, and the training process of the DQN algorithm is as follows:

[0144] T1: Randomly initialize Q network parameters θ and target Q network parameters θ' = θ, and create an experience replay pool D;

[0145] T2: Repeat the following steps until convergence:

[0146] T21: Get a sample (s, a, r, s') from the environment, where s is the current state, a is the action taken, r is the reward obtained, and s' is the next state;

[0147] T22: Add the sample (s, a, r, s') to the experience replay pool D;

[0148] T23: Randomly sample a batch of samples from the experience replay pool D;

[0149] T24: For each sample (s i , a i , r i , s′ i ), calculate the target value y i :

[0150] y i = ri +γmax a Q(s′ i ,a;θ′), where γ is the discount factor;

[0151] T25: Update the Q-network parameters θ using the mean squared error loss function:

[0152] Gradient descent can be used to minimize the loss function. Using gradient descent to minimize the loss function is a mature technique in the existing technology and a common practice in this field, so it will not be elaborated on here.

[0153] T26: Periodically copy the parameters of the Q network to the target Q network;

[0154] T3: After training, the trained Q network can be used to select actions. Given a state s, select action a that maximizes Q(s,a;θ).

[0155] After training, the learned Q-network can be deployed to the actual system, and the best action can be selected for peak shaving control based on real-time status monitoring results.

[0156] By utilizing reinforcement learning algorithms (such as DQN) to optimize peak-shaving control strategies, the output of each thermal power unit is adaptively adjusted based on real-time status monitoring results. This method combines factors such as frequency stability, output constraints, economy, and unit status to design a reward function, effectively balancing system safety and economy. The DQN algorithm improves learning efficiency through experience replay and target network mechanisms, achieving rapid convergence and continuous optimization. After training, the system can select the optimal action based on real-time data, significantly improving the stability, economy, and operating efficiency of the power grid.

[0157] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and its improved concept, should be covered within the scope of protection of the present invention.

Claims

1. A method for intelligent cruise control of peak load of thermal power unit based on big data analysis, characterized in that, The method comprises the following steps: Step 1: Real-time collection of operation data of the thermal power generating unit, and pre-processing of the collected operation data; Step 2: Collection of historical operation data of the thermal power generating unit, construction of a power grid load prediction model, and prediction of the power grid load in a future period of time through the model; Step 3: Real-time monitoring of key operation parameters of the thermal power generating unit, and monitoring and evaluation of the operation state of the thermal power generating unit; Step 4: Formulation of a corresponding peak regulation control strategy according to the power grid load prediction result and the thermal power generating unit state monitoring result; The specific process of Step 4 is as follows: The power grid load forecast results and thermal power unit condition monitoring results were obtained. The power grid load forecast results are the predicted power grid load values ​​for a future period of time. The status monitoring results of the thermal power unit indicate that the thermal power unit is in excellent, normal, or abnormal operating condition. Suppose there are N thermal power units, get the maximum output P max,i and the actual power grid load L st , the running state S i of the thermal power unit is represented by a numerical value: Si∈{1, 0.5, 0}, 1 represents excellent running state of thermal power unit, 0.5 represents normal running state of thermal power unit, 0 represents abnormal running state of thermal power unit, the running state S i of the thermal power unit is converted into weight W i : ； The content of the formulated peak regulation control strategy is the adjustment of the output distribution of each thermal power generating unit, and the initial output distribution is calculated using a weighted average method: Total weighted output capacity is calculated: ; Calculate the output proportion of each unit: ; Calculate the initial output of each unit: ; Then, a PI control algorithm is used to adjust the output of the thermal power generating unit, to compensate for the prediction error and meet the safety constraints: Prediction error computation: ; PI control algorithm: ; where K p and K i are proportional and integral gains, which need to be adjusted according to system characteristics and load prediction accuracy; Output adjustment: ; Safety constraints and output limits: ; where ΔP rate is the limit of the power output change rate of the thermal power unit, which represents the maximum value of the allowed change of the unit output per unit time, and Δt is the control period. Will be sent to each thermal power unit for control, adjust the output distribution of each thermal power unit, in each control cycle, re-acquire data, re-calculate W i , re-PI control, and adjust K p and K i according to the actual situation, real-time adjustment of thermal power unit output to track the actual load change; Step 5: Optimization of the peak regulation control strategy based on the thermal power generating unit state monitoring result, by using a reinforcement learning algorithm.

2. The method of claim 1, wherein, The specific process of Step 2 is as follows: The historical operation data of the thermal power generating unit is obtained, and the operation data includes the boiler pressure, the boiler temperature, the feed water flow, the turbine speed, the generator voltage, the generator current, the fuel flow, the power grid frequency, the power grid voltage, the load demand, the ambient temperature and the ambient humidity, a collection period is generated, and the collection period is equally divided into a plurality of collection time periods; The load demand change rate of the thermal power generating unit in the plurality of collection time periods is obtained, the load demand change rate represents the ratio between the load demand change amount and the corresponding time period length, a set A of load demand change rates is constructed in this way, and the average value of the difference between the maximum subset and the minimum subset in the set A is recorded as a load demand change rate difference FBC; The methods for calculating the load demand change rate difference FBC, the ambient temperature change rate difference WBC, the power grid frequency change rate difference PBC and the power grid voltage change rate difference DBC are the same; The load demand of the thermal power generating unit in the plurality of collection time periods is obtained, and the arithmetic average of the plurality of load demands is calculated, and the arithmetic average of the plurality of load demands is recorded as an average load demand PFX; The methods for calculating the average load demand PFX, the average ambient temperature PHW, the average power grid frequency PDP and the average power grid voltage PDY are the same.

3. The method of claim 2, wherein, The load demand change rate difference FBC, the ambient temperature change rate difference WBC, the power grid frequency change rate difference PBC, the power grid voltage change rate difference DBC, the average load demand PFX, the average ambient temperature PHW, the average power grid frequency PDP and the average power grid voltage PDY are combined to construct a power grid load prediction matrix DFJ, the power grid load prediction matrix DFJ is used as the input of a machine learning model, the power grid load in a future period of time corresponding to each group of power grid load prediction matrix DFJ is used as the output of the machine learning model, the power grid load in the future period of time is taken as the prediction target, the sum of the prediction errors of all training data is minimized as the training target, the machine learning model is trained until the sum of the prediction errors reaches convergence, and the power grid load prediction model is obtained. Real-time operation data of the thermal power generating unit is acquired, which is converted into a corresponding power grid load prediction matrix DFJ, which is input into a power grid load prediction model, and a power grid load prediction value in a future period of time is obtained through the power grid load prediction model , t = 1, 2, …, T, and T is a prediction step.

4. The method of claim 1, wherein, The specific process of the third step of monitoring and evaluating the operation state of the thermal power unit is as follows: Obtain the key operation parameters of the thermal power unit, including the boiler pressure, the turbine speed, the fuel flow and the load demand, generate a monitoring period, and equally divide the monitoring period into multiple monitoring time periods; Obtain the boiler pressure imbalance value of the thermal power unit in the monitoring period, and mark the boiler pressure imbalance value as GYS. The boiler pressure imbalance value represents the ratio between the part where the boiler pressure variation difference in each monitoring time period is greater than the preset boiler pressure variation difference threshold and the boiler pressure variation difference, and the boiler pressure variation difference represents the difference between the maximum value and the minimum value of the boiler pressure. The method for calculating the boiler pressure imbalance value GYS can be used to obtain the turbine speed imbalance value QZS, the fuel flow imbalance value RLS and the load demand imbalance value FXS.

5. The method of claim 4, wherein, After obtaining the boiler pressure imbalance value GYS, the turbine speed imbalance value QZS, the fuel flow imbalance value RLS and the load demand imbalance value FXS, the operation state evaluation coefficient YZP is calculated by the following formula: ； Wherein d1, d2, d3 and d4 are all preset proportional factor coefficients, and d4>d3>d2>d1>0. Compare the operation state evaluation coefficient YZP with the preset first operation state evaluation coefficient threshold and the preset second operation state evaluation coefficient threshold, and the preset first operation state evaluation coefficient threshold is less than the preset second operation state evaluation coefficient threshold: If the operation state evaluation coefficient YZP is less than the preset first operation state evaluation coefficient threshold, it indicates that the operation state of the thermal power unit is excellent; If the operation state evaluation coefficient YZP is greater than or equal to the preset first operation state evaluation coefficient threshold and less than the preset second operation state evaluation coefficient threshold, it indicates that the operation state of the thermal power unit is normal; If the operation state evaluation coefficient YZP is greater than or equal to the preset second operation state evaluation coefficient threshold, it indicates that the operation state of the thermal power unit is abnormal.

6. The method of claim 1, wherein, The specific process of the fifth step of optimizing the peak shaving control strategy by using the reinforcement learning algorithm is as follows: Obtain the thermal power unit state monitoring result, which is excellent, normal or abnormal, and the specific steps of optimizing the peak shaving control strategy by using the reinforcement learning algorithm are as follows: First, define a state space Z, which contains the following information: the operating state S of each thermal power unit i , the current grid load L st , the current output P of each thermal power unit i , the predicted value of the grid load in a future period of time The state space Z can be represented as a vector: ; Redefined action space A, action space A represents adjusting the output of each thermal power unit, and the action is represented as a vector: a=[∆P1,∆P2,…,∆P N ],where ∆P i represents the output adjustment of thermal power unit i.

7. The method of claim 6, wherein, Then define the reward function R, which considers multiple factors: Frequency stability: The reward function should encourage to keep the system frequency stable, using the square of the frequency deviation as a penalty term: R f = -K f (∆f) 2 where ∆f is the frequency deviation, k f is the weight coefficient; Output constraint: If the output of a unit exceeds the upper limit or is lower than the lower limit, a penalty is given: where P i,min and P i,max are the lower and upper limits of the power output of the unit i, respectively, 1 is an indicator function, and k c is a weight coefficient. Economy: Using fuel consumption rate as the penalty term, a model of the relationship between fuel consumption rate and unit output needs to be established. Assuming that the relationship is F i (P i ), then the penalty term is: where k e is a weight coefficient; Unit state: Preferably use the unit with excellent operation state, which can be rewarded: where k s is a weight coefficient; The total reward function can be represented as: R = R f + R c + R e + R s .

8. The method of claim 7, wherein, Finally, learn the optimal peak shaving control strategy using the DQN algorithm. The core of the DQN algorithm is to learn a Q network that approximates the Q function Q(s, a; θ), where θ is the network parameter. The training process of the DQN algorithm is as follows: T1: Randomly initialize the Q network parameter θ and the target Q network parameter θ'=θ, and create an experience replay pool D; T2: Repeat the following steps until convergence: T21: Obtain a sample (s, a, r, s') from the environment, where s is the current state, a is the action taken, r is the reward obtained, and s' is the next state; T22: Add the sample (s, a, r, s') to the experience replay pool D; T23: Randomly sample a batch of samples from the experience replay pool D; T24: for each sample , calculate the target value y i : where γ is a discount factor; T25: Update the Q-network parameters θ using the mean squared error loss function: Gradient descent can be used to minimize the loss function; T26: Periodically copy the parameters of the Q-network to the target Q-network; T3: After training, the trained Q-network can be used to select actions. Given a state s, select the action a that maximizes Q(s, a; θ); After training, the learned Q-network can be deployed to the actual system. According to the real-time state monitoring results, the best action is selected for peak shaving control.

Images