Novel power system generator set carbon emission management and control method and system
By constructing a predictive model of new energy generator sets and optimizing the output of traditional generator sets, the carbon emission optimization problem of the power system under the uncertainty of new energy generators is solved, and the efficient and economical carbon emission reduction goal is achieved.
Patent Information
- Application Number
- CN202510084876.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-01-20
- Publication Date
- 2025-04-25
AI Technical Summary
While ensuring the stability and economical power supply, it is difficult to optimize carbon emissions of generator sets, especially under the intermittent and uncertainty of new energy power generation.
By collecting equipment status data and weather data of power system generator sets, a prediction model of new energy generator sets is constructed, combined with LSTM and CNN for prediction, and using Monte Carlo simulation to quantify uncertainty. Based on the prediction results and power load requirements, the output of traditional generator sets is optimized, the output power and fuel usage are dynamically adjusted, and the optimal control of carbon emissions is achieved using DQN and strategic gradient methods.
Accurate prediction of the output power of new energy generator sets has been achieved, effectively reducing carbon emissions, improving the economic benefits and operating efficiency of the system, adapting to the uncertainty of new energy generation and optimizing the environmental pressure of traditional power generation.
Smart Images
Figure CN119887485A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of carbon emission management, and in particular to a novel method and system for controlling carbon emission of power generation units in a power system. Background Art
[0002] As one of the main sources of energy consumption and carbon emissions, the power industry is facing tremendous pressure to reduce emissions. At the same time, with the rapid development of renewable energy technology, the penetration rate of new energy such as wind power and solar energy in the power system continues to increase. However, due to the intermittent and uncertain nature of these new energy forms of power generation, traditional power dispatching and control technologies are difficult to meet the needs of contemporary power systems.
[0003] This poses a new challenge in power system management: how to optimize the carbon emissions of power generators while ensuring the stability and economy of power supply. Modern power systems not only need to introduce advanced prediction and optimization algorithms, but also need to use the Internet and Internet of Things technologies to achieve real-time data collection and processing. The development of these technologies has made it possible to solve complex scheduling problems in power systems. Summary of the invention
[0004] In order to solve the above problems, the purpose of the present invention is to provide a new carbon emission control method and system for power system generator sets, which can not only effectively reduce carbon emissions and help achieve environmental protection goals, but also improve the economic benefits and operating efficiency of the system.
[0005] To achieve the above object, the present invention adopts the following technical solutions:
[0006] A novel method for controlling carbon emissions of power generation units in a power system comprises the following steps:
[0007] S1: Collecting equipment status data and weather data of power system generator sets through sensors and IoT devices, wherein the generator sets include new energy generator sets and traditional generator sets;
[0008] S2: Preprocess the collected data, including removing outliers, filling missing data and data standardization;
[0009] S3: Build a prediction model for processing new energy generator sets, obtain the predicted output of new energy generator sets based on real-time data, and use Monte Carlo simulation to quantify uncertainty and obtain the predicted output confidence interval;
[0010] S4: Use the regression model to predict the power load demand in the corresponding time period;
[0011] S5: Based on the predicted output confidence interval of the new energy and the power load demand in the corresponding time period, the predicted output interval of the traditional generator set is obtained;
[0012] S6: Based on the predicted output of traditional generator sets, set constraints and optimization goals, use DQN and policy gradient methods to optimize the output of traditional generator sets, and dynamically adjust the output power and fuel usage of traditional generator sets to achieve optimal control of carbon emissions.
[0013] Furthermore, S2 is specifically:
[0014] Use Z-score to detect and remove outliers, and use the rule engine to set data thresholds and condition checks to automatically filter out abnormal data; set reasonable upper and lower thresholds based on historical data and expert knowledge, and data outside the range will be filtered out;
[0015] The missing value interpolation based on time series uses Lagrange interpolation to complete the data;
[0016] Data from different sources and dimensions are normalized to eliminate the differences between large and small dimensions.
[0017] Furthermore, the missing value interpolation based on time series uses Lagrange interpolation to complete the data, as follows:
[0018] According to the location of the missing value, several known data points near the missing value are selected as interpolation nodes;
[0019] Compute the Lagrangian basis function l for each known data point x j′ (x);
[0020]
[0021] Among them, x j′ is the value of the independent variable for the j′th known data point; x m′ For all except x j′ The value of the independent variable outside; n′ is the number of data points;
[0022] Combine all basis functions into an interpolation polynomial L(x);
[0023]
[0024] Among them, y j′ is the function value of the known data point;
[0025] Substitute the horizontal coordinate of the missing value into L(x) and calculate the completed value
[0026]
[0027] Among them, x missing is a missing value.
[0028] Furthermore, the prediction model for processing new energy generator sets is constructed by combining LSTM and CNN, including CNN, LSTM and fully connected layers, as follows:
[0029] The CNN layer extracts local temporal and spatial features of the data through convolution operations;
[0030]
[0031] Among them, X l-1 is the input of the previous layer, K l is the convolution kernel of layer l, b l is the bias, σ is the activation function; represents the sum of all covered points (m, n); (i, j) is the index of the point; S l (i,j) is the output of the convolutional layer of layer l;
[0032] Pooling operation is performed through the pooling layer:
[0033] P(i,j)=pool(S(i,j));
[0034] Among them, pool is the pooling function; P(i,j) is the feature after pooling, and S(i,j) is the output of the convolution layer;
[0035] Flatten the output of the last convolutional layer into a one-dimensional feature vector:
[0036] f CNN =f latten (P);
[0037] Among them, f CNN represents the output of CNN, f latten (P) indicates a flattening operation;
[0038] The feature vector extracted by CNN is used as the input sequence of LSTM:
[0039] x t =f CNN ;
[0040] LSTM unit calculation:
[0041] f t =σ(W f ·|h t-1 ,x t |+b f );
[0042] I t =σ(W I ·|h t-1 ,x t |+b I);
[0043]
[0044]
[0045] o t =σ(W o ·|h t-1 ,x t |+b0);
[0046] h t =o t tanh(C t );
[0047] Among them, x t is the input feature vector at time step t; h t-1 is the hidden state at time step t; f t For the Forget Gate; t is the input gate; is the candidate value, i.e. the candidate memory cell at the current moment; C t is the cell state, indicating the current state of the memory cell; C t-1 is the state of the memory cell at the previous moment; o t is the output gate; h t is the hidden state at the current moment; W f , W I , W C , W o are the weight matrices of the forget gate, input gate, candidate value, and output gate respectively; b f , b I , b C , b0 are the bias vectors of the forget gate, input gate, candidate value and output gate respectively; σ is the activation function;
[0048] The output of LSTM is passed to the fully connected layer to obtain the final power generation prediction;
[0049]
[0050] Among them, W out and b out are the weights and biases of the output layer;
[0051] During the training process, the Adam optimizer is used to update the model parameters to obtain the new energy generator set processing prediction model.
[0052] Furthermore, the Adam optimizer is used to update the model parameters to obtain the new energy generator processing prediction model, as follows:
[0053] Initialize parameters, including learning rate η, momentum factors β1, β2, and constant ∈, initialize momentum vector m0 = 0, v0 = 0; in each training iteration:
[0054] Calculate the model output and loss function value L;
[0055] Calculate the gradient of the loss function L for the model parameters θ:
[0056]
[0057] Among them, L(θ) is the loss function with model parameter θ;
[0058] Compute the gradient norm:
[0059]
[0060] Among them, g t,i′ is the gradient vector g t The i′th component of
[0061] If g t ≥τ, perform gradient clipping:
[0062]
[0063] τ
[0064] Among them, is the gradient clipping threshold;
[0065] First-order momentum update:
[0066] m t =β1·m t-1 +(1-β1)·gt;
[0067] Among them, m t is the updated first-order momentum; m t-1 is the current first-order momentum; the second-order momentum is updated:
[0068]
[0069] Among them, v t is the updated second-order momentum; v t-1 is the current first-order momentum; first-order momentum correction:
[0070]
[0071] in, is the corrected first-order momentum;
[0072] Second-order momentum correction:
[0073]
[0074] in, is the corrected second-order momentum;
[0075] Use the Adam formula to update the parameters:
[0076]
[0077] Among them, θ t+1 is the updated parameter; θ t is the current parameter;
[0078] The training cycle continues until the loss function value is less than the threshold or reaches the maximum number of iterations.
[0079] Furthermore, Monte Carlo simulation is used to quantify uncertainty and obtain the predicted output confidence interval, as follows:
[0080] Identify sources of uncertainty that affect forecast results, including: uncertainty in input variables, such as wind speed and sunlight intensity; inaccuracy in model parameters;
[0081] Select the corresponding probability distribution for each uncertainty factor. For wind speed, use the normal distribution. For bounded data, for light intensity, use the uniform distribution or triangular distribution.
[0082] Use probability distributions to generate multiple samples for each uncertain variable;
[0083] For each set of samples, calculate the output of the prediction model and record the results:
[0084]
[0085] Among them, x i″′ is the input vector of the i′′th sampling, is the output of the prediction model; χ is the introduced random error;
[0086] Among them, x i is the input vector of the i-th sampling; χ is the introduced random error;
[0087] Based on the obtained set of prediction results, calculate the standard deviation;
[0088] Calculate the confidence interval CI of the predicted output with a given confidence level:
[0089]
[0090] Among them, y is the mean of the prediction results and σ′ is the standard deviation.
[0091] Furthermore, the regression model is based on SVR and uses the RBF kernel function:
[0092] K(x i″ ,x j″ ) = exp(-γ||x i″ -x j″ || 2 );
[0093] Among them, γ is the support vector factor; x i″ and x j″ are the i″th and j″th input variables respectively;
[0094] Set SVR parameters, including penalty coefficient C and regression accuracy o;
[0095] Train the SVR model, and the optimization goal is to minimize the loss function:
[0096]
[0097] Among them, w is the weight vector, b is the bias term, ξ i″ , is the slack variable corresponding to the i″th input variable; n″ is the number of input variables;
[0098] Constraints:
[0099] y i″ -(w·φ(x i″ )+b)≤o+ξ i″ ;
[0100]
[0101]
[0102] Among them, φ(x i″ ) is the input data x i″ Mapping in high-dimensional feature space; y i″ is the actual observed value;
[0103] Use grid search or random search to determine the best C, o, and γ parameter combination;
[0104] Use the trained SVR model to make predictions for new time periods.
[0105] Furthermore, S6 is specifically:
[0106] Define the state space S, including the current output power of the generator set, the current fuel usage, the current weather data, the load demand, the market price and the current carbon emission level;
[0107] Define action space A, including adjusting the output power of the generator set; switching the fuel type; starting or stopping the generator set operation;
[0108] Define the reward function R, considering the comprehensive benefits of economy and environmental protection:
[0109] R=-(C oper +λ′·C emission )
[0110] Among them, C oper is the operating cost, C emission is the emission cost, λ′ is the coefficient for weighing the economic and environmental protection;
[0111] Constraints include meeting grid requirements for power supply stability and load demand, and emissions must not exceed preset thresholds;
[0112] Use a neural network to approximate the Q-value function θ′, where s is the state, a is the action, and θ′ is the network parameter;
[0113] Update the Q value using the modified Bellman equation update rule:
[0114]
[0115] Among them, α is the learning rate, γ′ is the discount factor, is the target network parameter, r is the immediate reward of the current step;
[0116] Use the experience replay pool to sample experience batches for Q value update;
[0117] And use the policy gradient method to select actions;
[0118] The obtained strategy is applied to the real-time system to make the optimal decision using the current state.
[0119] Furthermore, the policy gradient method is used to select actions, as follows:
[0120] Set up a policy network π(s,a;θ′) to output the probability distribution of actions;
[0121] Optimize parameters using the policy gradient method:
[0122]
[0123] Among them, the performance measure of the J(θ′) strategy is usually the expectation of the cumulative reward; E[.] is the expectation function;
[0124] Implement the Actor-Critic framework and introduce the baseline function V π (s) to reduce variance:
[0125]
[0126] A novel carbon emission control system for power system generator sets includes a processor, a memory, and a computer program stored in the memory. When the processor executes the computer program, it specifically executes the steps in the novel carbon emission control method for power system generator sets as described above.
[0127] The present invention has the following beneficial effects:
[0128] 1. The present invention can more flexibly and adaptably cope with the uncertainty of new energy power generation and the environmental pressure brought by traditional power generation, and effectively achieve the maximum utilization of clean energy and carbon emission reduction goals;
[0129] 2. The present invention combines LSTM and CNN to construct a new energy generator processing prediction model, uses the feature extraction ability of the convolutional layer and the sequence learning ability of LSTM to achieve accurate prediction of the output power of the new energy generator, and uses the Adam optimization algorithm to accelerate convergence. The improved LSTM-CNN hybrid model can more effectively and robustly predict the output power of the new energy generator, and combines gradient clipping and Nesterov momentum to effectively prevent gradient explosion and use momentum to accelerate convergence;
[0130] 3. The present invention effectively uses DQN and the policy gradient method in the output control of the generator set to reasonably adjust the power and fuel usage to achieve optimal control of carbon emissions. BRIEF DESCRIPTION OF THE DRAWINGS
[0131] Figure 1 The figure is a flow chart of the method of the present invention. DETAILED DESCRIPTION
[0132] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments:
[0133] refer to Figure 1 In this embodiment, a novel method for controlling carbon emissions of power generation units in a power system includes the following steps:
[0134] S1: Collecting equipment status data and weather data of power system generator sets through sensors and IoT devices, wherein the generator sets include new energy generator sets and traditional generator sets;
[0135] S2: Preprocess the collected data, including removing outliers, filling missing data and data standardization;
[0136] S3: Build a prediction model for processing new energy generator sets, obtain the predicted output of new energy generator sets based on real-time data, and use Monte Carlo simulation to quantify uncertainty and obtain the predicted output confidence interval;
[0137] S4: Use the regression model to predict the power load demand in the corresponding time period;
[0138] S5: Based on the predicted output confidence interval of new energy (wind energy, solar energy) and the power load demand in the corresponding time period, obtain the predicted output interval of the traditional generator set;
[0139] S6: Based on the predicted output of traditional generator sets, set constraints (such as emission reduction targets, economic efficiency, etc.) and optimization goals (such as minimizing emissions or fuel usage), use Deep Q-Networks (DQN) and policy gradient methods to optimize the output of traditional generator sets, and dynamically adjust the output power and fuel usage of traditional generator sets to achieve optimal control of carbon emissions.
[0140] In this embodiment, S2 is specifically:
[0141] Use Z-score to detect and remove outliers, and use rule engines (such as Drools) to set data thresholds and condition checks to automatically filter out abnormal data; set reasonable upper and lower thresholds based on historical data and expert knowledge, and data that exceeds the range is filtered out;
[0142] The missing value interpolation based on time series uses Lagrange interpolation to complete the data;
[0143] Data from different sources and dimensions are normalized to eliminate the differences between large and small dimensions.
[0144] In this embodiment, the missing value interpolation based on the time series adopts Lagrange interpolation to complete the data, as follows:
[0145] According to the location of the missing value, select several known data points near the missing value (such as 2-3 points before and after each) as interpolation nodes;
[0146] Compute the Lagrangian basis function l for each known data point x j′ (x);
[0147]
[0148] Among them, x j′ is the value of the independent variable for the j′th known data point; x m′ For all except x j′ The value of the independent variable outside; n′ is the number of data points;
[0149] Combine all basis functions into an interpolation polynomial L(x);
[0150]
[0151] Among them, y j′is the function value of the known data point;
[0152] Substitute the horizontal coordinate of the missing value into L(x) and calculate the completed value :
[0153]
[0154] Among them, x missing is a missing value.
[0155] In this embodiment, the new energy generator processing prediction model is constructed by combining LSTM and CNN, including CNN, LSTM and fully connected layers, as follows:
[0156] The CNN layer extracts local temporal and spatial features of the data through convolution operations;
[0157]
[0158] Among them, X l-1 is the input of the previous layer, K l is the convolution kernel of layer l, b l is the bias, σ is the activation function; represents the sum of all covered points (m, n), (i, j) is the index of the point; S l (i,j) is the output of the convolutional layer of layer l;
[0159] Pooling operation is performed through the pooling layer:
[0160] P(i,j)=pool(S(i,j));
[0161] Among them, pool is the pooling function; P(i,j) is the feature after pooling, and S(i,j) is the output of the convolution layer;
[0162] Flatten the output of the last convolutional layer into a one-dimensional feature vector:
[0163] f CNN =f latten (P);
[0164] Among them, f CNN represents the output of CNN, f latten (P) indicates a flattening operation;
[0165] The feature vector extracted by CNN is used as the input sequence of LSTM:
[0166] x t =f CNN
[0167] LSTM unit calculation:
[0168] f t =σ(W f ·|h t-1 ,x t |+b f );
[0169] I t =σ(W I ·|h t-1 ,x t |+b I );
[0170]
[0171]
[0172] o t =σ(W o ·|h t-1 ,x t |+b0);
[0173] h t =o t tanh(C t );
[0174] Among them, x t is the input feature vector at time step t; h t-1 is the hidden state at time step t; f t For the Forget Gate; t is the input gate; is the candidate value, i.e. the candidate memory cell at the current moment; C t is the cell state, indicating the current state of the memory cell; C t-1 is the state of the memory cell at the previous moment; o t is the output gate; h t is the hidden state at the current moment; W f , W I , W C , W o are the weight matrices of the forget gate, input gate, candidate value, and output gate respectively; b f , b I , b C , b0 are the bias vectors of the forget gate, input gate, candidate value and output gate respectively; σ is the activation function;
[0175] The output of LSTM is passed to the fully connected layer to obtain the final power generation prediction;
[0176]
[0177] Among them, W out and b outare the weights and biases of the output layer;
[0178] During the training process, the Adam optimizer is used to update the model parameters to obtain the new energy generator set processing prediction model.
[0179] In this embodiment, the Adam optimizer is used to update the model parameters to obtain the new energy generator processing prediction model, which is as follows:
[0180] Initialize parameters, including learning rate η, momentum factors β1, β2, and constant ∈, initialize momentum vector m0=0, v0=0;
[0181] In each training iteration:
[0182] Calculate the model output and loss function value L (using mean square error MSE);
[0183] Calculate the gradient of the loss function L for the model parameters θ:
[0184]
[0185] Among them, L(θ) is the loss function with model parameter θ;
[0186] Compute the gradient norm:
[0187]
[0188] Among them, g t，i′ is the i′th component of the gradient vector gt;
[0189] If g t ≥τ, perform gradient clipping:
[0190]
[0191] Among them, τ is the gradient clipping threshold;
[0192] First-order momentum update:
[0193] m t =β1·m t-1 +(1-β1)·g t ;
[0194] Among them, m t is the updated first-order momentum; m t-1 is the current first-order momentum;
[0195] Second-order momentum update:
[0196]
[0197] Among them, v tis the updated second-order momentum; v t-1 is the current first-order momentum;
[0198] First-order momentum correction:
[0199]
[0200] in, is the corrected first-order momentum;
[0201] Second-order momentum correction:
[0202]
[0203] in, is the corrected second-order momentum;
[0204] Use the Adam formula to update the parameters:
[0205]
[0206] Among them, θ t+1 is the updated parameter; θ t is the current parameter;
[0207] The training cycle continues until the loss function value is less than the threshold or reaches the maximum number of iterations.
[0208] In this embodiment, Monte Carlo simulation is used to quantify uncertainty and obtain the predicted output confidence interval, as follows:
[0209] Identify sources of uncertainty that affect forecast results, including: uncertainty in input variables, such as wind speed and sunlight intensity; inaccuracy in model parameters;
[0210] Select the corresponding probability distribution for each uncertainty factor. For wind speed, use the normal distribution. For bounded data, for light intensity, use the uniform distribution or triangular distribution.
[0211] Use probability distributions to generate multiple samples for each uncertain variable;
[0212] For each set of samples, calculate the output of the prediction model and record the results:
[0213]
[0214] Among them, x i″′ is the input vector of the i′′th sampling, is the output of the prediction model; χ is the introduced random error;
[0215] Based on the obtained set of prediction results, calculate the standard deviation;
[0216] Calculate the confidence interval CI of the predicted output with a given confidence level (such as 95%):
[0217]
[0218] in, is the mean of the prediction results, and σ′ is the standard deviation.
[0219] In this embodiment, the regression model is based on SVR and uses the RBF kernel function:
[0220] K(x i″ ,x j″ ) = exp(-γ||x i″ -x j″ || 2 );
[0221] Among them, γ is the support vector factor; x i″ and x j″ are the i″th and j″th input variables respectively;
[0222] Set SVR parameters, including penalty coefficient C and regression accuracy o;
[0223] Train the SVR model, and the optimization goal is to minimize the loss function:
[0224]
[0225] Among them, w is the weight vector, b is the bias term, ξ i″ , is the slack variable corresponding to the i″th input variable; n″ is the number of input variables;
[0226] Constraints:
[0227] y i″ -(w·φ(x i″ )+b)≤o+ξ i″ ;
[0228]
[0229]
[0230] Among them, φ(x i″ ) is the input data x i″ Mapping in high-dimensional feature space; y i″ is the actual observed value;
[0231] Use grid search or random search to determine the best C, o, and γ parameter combination;
[0232] Use the trained SVR model to make predictions for new time periods.
[0233] In this embodiment, S6 is specifically:
[0234] Define the state space S, including the output power of the current generator set, the current fuel usage (such as fuel type and its consumption), current weather data, load demand, market price and current carbon emission level;
[0235] Define action space A, including adjusting the output power of the generator set; switching the fuel type; starting or stopping the generator set operation;
[0236] Define the reward function R, considering the comprehensive benefits of economy and environmental protection:
[0237] R=-(C oper +λ′·C emission )
[0238] Among them, C oper is the operating cost, C emission is the emission cost, λ′ is the coefficient for weighing the economic and environmental protection;
[0239] Constraints include meeting grid requirements for power supply stability and load demand, and emissions must not exceed preset thresholds;
[0240] Use a neural network to approximate the Q-value function θ′, where s is the state, a is the action, and θ′ is the network parameter;
[0241] Update the Q value using the modified Bellman equation update rule:
[0242]
[0243] Among them, α is the learning rate, γ′ is the discount factor, is the target network parameter, r is the immediate reward of the current step;
[0244] Use the experience replay pool to sample experience batches for Q value update;
[0245] And use the policy gradient method to select actions;
[0246] The obtained strategy is applied to the real-time system to make the optimal decision using the current state.
[0247] In this embodiment, the policy gradient method is used to select actions, as follows:
[0248] Set up a policy network π(s,a;θ′) to output the probability distribution of actions;
[0249] Optimize parameters using the policy gradient method:
[0250]
[0251] Among them, the performance measure of the J(θ′) strategy is usually the expectation of the cumulative reward; E[.] is the expectation function;
[0252] Implement the Actor-Critic framework and introduce the baseline function V π (s) to reduce variance:
[0253]
[0254] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0255] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.
[0256] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.
[0257] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1The steps for the functions specified in one or more boxes.
[0258] The above is only a preferred embodiment of the present invention, and does not limit the present invention in other forms. Any technician familiar with the profession may use the above disclosed technical content to change or modify it into an equivalent embodiment with equivalent changes. However, any simple modification, equivalent change and modification made to the above embodiment according to the technical essence of the present invention without departing from the technical solution of the present invention still belongs to the protection scope of the technical solution of the present invention.
Claims
1. A novel method for controlling carbon emissions of power generation units in a power system, characterized in that: The following steps are involved: S1: Collecting equipment status data and weather data of power system generator sets through sensors and IoT devices, wherein the generator sets include new energy generator sets and traditional generator sets; S2: Preprocess the collected data, including removing outliers, filling missing data and data standardization; S3: Build a prediction model for processing new energy generator sets, obtain the predicted output of new energy generator sets based on real-time data, and use Monte Carlo simulation to quantify uncertainty and obtain the predicted output confidence interval; S4: Use the regression model to predict the power load demand in the corresponding time period; S5: Based on the predicted output confidence interval of the new energy and the power load demand in the corresponding time period, the predicted output interval of the traditional generator set is obtained; S6: Based on the predicted output of traditional generator sets, set constraints and optimization goals, use DQN and policy gradient methods to optimize the output of traditional generator sets, and dynamically adjust the output power and fuel usage of traditional generator sets to achieve optimal control of carbon emissions.
2. A novel method for controlling carbon emissions of power generation units in a power system according to claim 1, characterized in that: The S2 is specifically: Use Z-score to detect and remove outliers, and use the rule engine to set data thresholds and condition checks to automatically filter out abnormal data; set reasonable upper and lower thresholds based on historical data and expert knowledge, and data outside the range will be filtered out; The missing value interpolation based on time series uses Lagrange interpolation to complete the data; Data from different sources and dimensions are normalized to eliminate the differences between large and small dimensions.
3. According to a novel carbon emission control method for power system generator sets as described in claim 2, it is characterized in that: The missing value interpolation based on time series adopts Lagrange interpolation to complete the data, as follows: According to the location of the missing value, several known data points near the missing value are selected as interpolation nodes; Compute the Lagrangian basis function l for each known data point x j′ (x); Among them, x j′ is the value of the independent variable for the j′th known data point; x m′ For all except x j′ The value of the independent variable outside; n′ is the number of data points; Combine all basis functions into an interpolation polynomial L(x); Among them, y j′ is the function value of the known data point; Substitute the horizontal coordinate of the missing value into L(x) and calculate the completed value Among them, x missing is a missing value.
4. A novel method for controlling carbon emissions of power generation units in a power system according to claim 1, characterized in that: The new energy generator processing prediction model is constructed by combining LSTM and CNN, including CNN, LSTM and fully connected layers, as follows: The CNN layer extracts local temporal and spatial features of the data through convolution operations; Among them, X l-1 is the input of the previous layer, K l is the convolution kernel of layer l, b l is the bias, σ is the activation function; represents the sum of all covered points (m, n); (i, j) is the index of the point; S l (i,j) is the output of the convolutional layer of layer l; Pooling operation is performed through the pooling layer: P(i,j)=pool(S(i,j)); Among them, pool is the pooling function; P(i,j) is the feature after pooling, and S(i,j) is the output of the convolution layer; Flatten the output of the last convolutional layer into a one-dimensional feature vector: f CNN =f latten (P); Among them, f CNN represents the output of CNN, f latten (P) indicates a flattening operation; The feature vector extracted by CNN is used as the input sequence of LSTM: x t =f CNN ; LSTM unit calculation: f t =σ(W f ·|h t-1 ,x t |+b f ); I t =σ(W I ·|h t-1 ,x t |+b I ); o t =σ(W o ·|h t-1 ,x t |+b0); h t =o t fishy(C) t ); Among them, x t is the input feature vector at time step t; h t-1 is the hidden state at time step t; f t For the Forget Gate; t is the input gate; is the candidate value, i.e. the candidate memory cell at the current moment; C t is the cell state, indicating the current state of the memory cell; C t-1 is the state of the memory cell at the previous moment; o t is the output gate; h t is the hidden state at the current moment; W f , W I , W C , W o are the weight matrices of the forget gate, input gate, candidate value, and output gate respectively; b f 、b I 、b C , b0 are the bias vectors of the forget gate, input gate, candidate value and output gate respectively; σ is the activation function; The output of LSTM is passed to the fully connected layer to obtain the final power generation prediction; Among them, W out and b out are the weights and biases of the output layer; During the training process, the Adam optimizer is used to update the model parameters to obtain the new energy generator set processing prediction model.
5. A novel method for controlling carbon emissions of power generation units in a power system according to claim 4, characterized in that: The Adam optimizer is used to update the model parameters to obtain the new energy generator processing prediction model, which is as follows: Initialize parameters, including learning rate η, momentum factors β1, β2, and constant ∈, and initialize momentum vector m0=0, v0=0; In each training iteration: Calculate the model output and loss function value L; Calculate the gradient of the loss function L for the model parameters θ: Among them, L(θ) is the loss function with model parameter θ; Compute the gradient norm: Among them, g t,i′ is the gradient vector g t The i′th component of If g t ≥τ, perform gradient clipping: Among them, τ is the gradient clipping threshold; First-order momentum update: m t =β1·m t-1 +(1-β1)·g t ; Among them, m t is the updated first-order momentum; m t-1 is the current first-order momentum; Second-order momentum update: Among them, v t is the updated second-order momentum; v t-1 is the current first-order momentum; First-order momentum correction: in, is the corrected first-order momentum; Second-order momentum correction: in, is the corrected second-order momentum; Use the Adam formula to update the parameters: Among them, θ t+1 is the updated parameter; θ t is the current parameter; The training cycle continues until the loss function value is less than the threshold or reaches the maximum number of iterations.
6. A novel method for controlling carbon emissions of power generation units in a power system according to claim 4, characterized in that: The Monte Carlo simulation is used to quantify uncertainty and obtain the predicted output confidence interval, as follows: Identify sources of uncertainty that affect forecast results, including: uncertainty in input variables, such as wind speed and sunlight intensity; inaccuracy in model parameters; Select the corresponding probability distribution for each uncertainty factor. For wind speed, use the normal distribution. For bounded data, for light intensity, use the uniform distribution or triangular distribution. Use probability distributions to generate multiple samples for each uncertain variable; For each set of samples, calculate the output of the prediction model and record the results: Among them, x i″′ is the input vector of the i′′th sampling, is the output of the prediction model; χ is the introduced random error; Based on the obtained set of prediction results, calculate the standard deviation; Calculate the confidence interval CI of the predicted output with a given confidence level: in, is the mean of the prediction results, and σ′ is the standard deviation.
7. A novel method for controlling carbon emissions of power generation units in a power system according to claim 1, characterized in that: The regression model is based on SVR and uses the RBF kernel function: K(x i″ ,x j″ )=exp(-γ||x i″ -x j″ || 2 ); Among them, γ is the support vector factor; x i″ and x j″ are the i″th and j″th input variables respectively; Set SVR parameters, including penalty coefficient C and regression accuracy o; Train the SVR model, and the optimization goal is to minimize the loss function: Among them, w is the weight vector, b is the bias term, is the slack variable corresponding to the i″th input variable; n″ is the number of input variables; Constraints: y i″ -(w·φ(x i″ )+b)≤o+ξ i″ ; Among them, φ(x i″ ) is the input data x i″ Mapping in high-dimensional feature space; y i″ is the actual observed value; Use grid search or random search to determine the best C, o, and γ parameter combination; Use the trained SVR model to make predictions for new time periods.
8. A novel method for controlling carbon emissions of power generation units in a power system according to claim 1, characterized in that: The S6 is specifically: Define the state space S, including the current output power of the generator set, the current fuel usage, the current weather data, the load demand, the market price and the current carbon emission level; Define action space A, including adjusting the output power of the generator set; switching the fuel type; starting or stopping the generator set operation; Define the reward function R, considering the comprehensive benefits of economy and environmental protection: R=-(C oper +λ′·C emission ); Among them, C oper is the operating cost, C emission is the emission cost, λ′ is the coefficient for weighing the economic and environmental protection; Constraints include meeting grid requirements for power supply stability and load demand, and emissions must not exceed preset thresholds; Use a neural network to approximate the Q-value function θ′, where s is the state, a is the action, and θ′ is the network parameter; Update the Q value using the modified Bellman equation update rule: Among them, α is the learning rate, γ′ is the discount factor, is the target network parameter, r is the immediate reward of the current step; Use the experience replay pool to sample experience batches for Q value update; And use the policy gradient method to select actions; The obtained strategy is applied to the real-time system to make the optimal decision using the current state.
9. A novel method for controlling carbon emissions of power generation units in a power system according to claim 8, characterized in that: The policy gradient method is used to select actions, as follows: Set up a policy network π(s,a;θ′) to output the probability distribution of actions; Optimize parameters using the policy gradient method: Among them, the performance measure of the J(θ′) strategy is usually the expectation of the cumulative reward; E[.] is the expectation function; Implement the Actor-Critic framework and introduce the baseline function V π (s) to reduce variance:
10. A new type of carbon emission control system for power system generator sets, characterized in that: It includes a processor, a memory and a computer program stored in the memory. When the processor executes the computer program, it specifically executes the steps in a new method for carbon emission control of power system generator sets as described in any one of claims 1-9.