Glass factory load low-carbon regulation and control method considering capacity compensation and carbon quota

Abstract

The invention discloses a glass factory load low-carbon regulation and control method considering capacity compensation and carbon quota, and aims to realize refined carbon emission control and flexible scheduling in the production process of a glass factory. By fusing multi-dimensional input of historical load power, load change rate, environment temperature, humidity, wind speed, electricity price, time and the like and utilizing an STFormer architecture to model a complex time dependency relationship, a kilowatt-hour electricity carbon emission factor in a future time period is predicted, and prior information is provided for low-carbon scheduling; with minimization of carbon emission and electric charge cost and maximization of capacity income as targets, a future load scheduling strategy is optimized under constraints of carbon quota, an electric power regulation range and equipment operation; and globally searching an optimal load control scheme through a quantum particle swarm algorithm, and generating a multi-scene-oriented dynamic regulation and control strategy under the constraints of a multi-target fitness function and a penalty mechanism. The method not only improves the precision and robustness of carbon factor prediction, but also realizes flexible scheduling of the glass factory load driven by carbon economy.

Description

Technical Field

[0001] The present invention relates to the technical field of power system regulation, and in particular to a low-carbon regulation method for glass factory loads taking into account capacity compensation and carbon quota. Background Art

[0002] Low-carbon load regulation in glass factories is crucial. On the one hand, it reduces the number of energy-intensive processes operating during periods of high grid carbon emissions, thereby increasing the utilization rate of green electricity. On the other hand, it helps achieve dynamic optimization of carbon emissions throughout the entire glass manufacturing process. Furthermore, low-carbon regulation can reduce carbon emission costs for companies, enhancing their market competitiveness and their image as socially responsible.

[0003] However, the existing load control methods for glass factories have the following shortcomings in achieving low-carbon goals: 1) Lack of high-precision carbon factor prediction capabilities: Current methods often rely on linear regression or static mean to predict the carbon emission factor per kilowatt-hour, which makes it difficult to capture the nonlinear dynamic characteristics of the carbon factor as it changes with time, power structure, and electricity load, resulting in lags or failures in the control strategy. 2) Insufficient consideration of capacity compensation and carbon quota mechanisms: Most control models ignore the coupling relationship between capacity compensation incentives and carbon quota constraints, and are unable to minimize carbon emissions while ensuring production continuity, limiting the room for improvement in control benefits. 3) Insufficient control strategy optimization capabilities: Traditional optimization strategies such as genetic algorithms or PSO converge slowly and are prone to falling into local optimality, making it difficult to achieve multi-objective, strongly coupled low-carbon load control in glass factories under the dual conditions of complex process constraints and carbon policies.

[0004] To this end, we designed a low-carbon load control method for glass plants that takes into account capacity compensation and carbon quotas to solve the above problems. Summary of the Invention

[0005] The purpose of the present invention is to solve the shortcomings of existing glass factory load control methods in achieving low-carbon goals, such as the lack of high-precision carbon factor prediction capability, insufficient consideration of capacity compensation and carbon quota mechanisms, and insufficient control strategy optimization capabilities. A low-carbon control method for glass factory loads considering capacity compensation and carbon quotas is proposed, and a per-kilowatt-hour carbon emission factor prediction model based on the STFormer architecture is proposed, which integrates time series modeling and spatial perception capabilities to accurately predict the dynamic changes of carbon factors; a low-carbon control model for glass factory loads taking into account capacity compensation and carbon quotas is constructed to achieve the unity of economy and low carbon; a low-carbon control strategy based on QPSO (quantum particle swarm optimization) is designed to improve control efficiency and global optimization capabilities, and support the green transformation of glass factories.

[0006] In order to achieve the above object, the present invention adopts the following technical solutions:

[0007] A low-carbon load control method for a glass plant considering capacity compensation and carbon quotas comprises the following steps:

[0008] Step 1: Construct a multi-dimensional input feature vector that integrates the glass factory's load characteristics, meteorological environment, electricity price, and time information. Use the STFormer architecture to predict the carbon emission factors per kilowatt-hour at multiple future moments, extract the temporal dependencies, and output the carbon emission factors per kilowatt-hour within the future window through a decoder.

[0009] Step 2: Based on the carbon emission factor per kilowatt-hour, combined with electricity price information, capacity compensation price, and equipment operation boundary parameters, a multi-objective low-carbon regulation optimization model is constructed;

[0010] Step three: For different operating scenarios, the quantum particle swarm optimization algorithm is introduced to construct a dynamic control strategy optimization model under multiple scenarios. The optimal load scheduling sequence is encoded as the particle position, and a multi-objective fitness function is defined to balance carbon emissions, electricity price costs and capacity benefits. The scheduling constraints are handled by introducing a penalty function. The quantum particle swarm optimization algorithm is run separately in multiple scenarios to obtain a set of optimal control strategies suitable for different carbon economic scenarios.

[0011] Further preferably, in step 1, constructing a multi-dimensional input feature vector integrating the glass factory load characteristics, meteorological environment, electricity price and time information includes the following steps:

[0012] Construct a multi-dimensional input feature vector and a historical feature sequence input sample as follows:

[0013]

[0014] X t-T+1:t =[x t-T+1 ,x t-T+2 ,...,x t ]

[0015] Where x t 、x t-T+2 and x t-T+1 Represents the input feature vector at different times, P t is the active power of the glass factory load at time t; is the load change rate at time t; T t is the ambient temperature at time t; H t is the humidity at time t; W t is the wind speed at time t; is the electricity price at time t; hour is the current hour; t day is the current day of the week; X t-T+1:t Represents the historical feature sequence input sample.

[0016] More preferably, the multi-dimensional input feature vector is injected into the historical feature sequence input sample, and position encoding is introduced, as shown in the following formula:

[0017]

[0018] The historical feature sequence is input into the sample input position encoding, as shown in the following formula:

[0019]

[0020] In the above formula, e τ is the position encoding vector at time τ, d is the embedding dimension, i∈[1,d / 2], is the position-encoded input sequence, E t-T+1:t is the position code sequence from time t-T+1 to time t, E t-T+1:t =[e t-T+1 ,e t-T+2 ,...,e t ];

[0021] Using the sparse attention mechanism, the dependencies between important features in the input sequence are calculated and the input sequence is encoded:

[0022]

[0023] in,

[0024] W Q , W K , W V is the corresponding weight matrix, d k is each vector dimension of Q and K after conversion. Attention is a calculation function for extracting the temporal variable dependency between each moment in the input sequence. H enc Represents the higher order of each moment; is the position-encoded input sequence, F STEnc (·) is the STFormer sequence encoding function, and the superscript T represents the transpose of the matrix;

[0025] By inputting the feature representation extracted by the encoder into the decoder, the carbon emission factor sequence of each kilowatt-hour in the future time window is predicted. Each moment of the output corresponds to an estimated carbon factor value, which describes the carbon emission level of the glass factory under the future operating state.

[0026] Generate future forecasts using the historically encoded state:

[0027]

[0028] Where, is the carbon emission factor per kilowatt-hour of electricity predicted at a future time, H is the future predicted scheduling duration; F STDec (·) is the STFormer decoder.

[0029] Further preferably, in step 2, based on the carbon emission factor per kilowatt-hour, combined with electricity price information, capacity compensation price, and equipment operation boundary parameters, a multi-objective low-carbon regulation optimization model is constructed:

[0030]

[0031] Where τ is a time point in the prediction window, t is the current time; H is the future prediction scheduling duration; P τ is the optimal dispatch load of the glass factory at time τ, is the predicted carbon emission factor per kilowatt-hour at time τ; is the electricity price at time τ; is the capacity compensation price at time τ; is the load reduction amplitude, which represents the capacity gain obtained from load reduction, where is the original planned operating load of the glass factory at time τ; α, β, γ are the weight parameters for controlling carbon emissions, electricity costs, and capacity benefits, respectively.

[0032] Furthermore, preferably, the multi-objective low-carbon control optimization model is constrained by the total carbon quota constraint, the upper and lower limit constraints of load scheduling, the load adjustment range constraint, and the minimum production line energy consumption constraint of the glass factory. The constraint formula is as follows:

[0033]

[0034] In the above formula, Δt is the duration of a single scheduling cycle, C quota is the total amount of carbon emission quota allowed in the current stage, are the minimum load boundary and maximum load boundary allowed by the equipment respectively; R max P is the maximum load change rate per hour; τ-1 is the dispatching load at the previous moment; E min The minimum total electricity required to maintain basic production capacity for a glass factory.

[0035] Further preferably, in step three, for different operating scenarios, the quantum particle swarm optimization algorithm is introduced to construct a dynamic control strategy optimization model under multiple scenarios.

[0036] Introducing quantum particle swarm optimization algorithm, particle position The scheduling of the i-th particle controls the load in the future H period, which represents the control strategy of the glass factory within the prediction window, that is, the load scheduling sequence. When initializing the population, a disturbance is added, as shown in the following formula:

[0037]

[0038] Where, is the dispatch load value of the i-th particle at time τ; The optimized scheduling value obtained by the multi-objective low-carbon regulation optimization model is used as a reference for particle initialization; is the initial disturbance term, satisfying ε~U(-δ,δ), U(-δ,δ) is the disturbance range, and δ is the disturbance threshold;

[0039] The dynamic control strategy optimization model constructed in multiple scenarios is as follows:

[0040]

[0041] Where τ is a time point within the prediction window; t is the current time; H is the future prediction scheduling duration; P τ is the optimal dispatch load of the glass factory at time τ; To predict the carbon emission factor per kilowatt-hour; is the electricity price at time τ; is the capacity compensation price at time τ; is the load reduction amplitude, which represents the capacity gain obtained from load reduction, where is the original planned operating load of the glass factory at time τ; (α i ,β i ,γ i ) is a weight combination, indicating that in three typical scenarios S i The weights under, where i = 1, 2, 3.

[0042] Further preferably, in step 3, three typical scenarios S are set for different operation scenarios. i To cope with different actual operating constraints:

[0043] In the S1 scenario, the capacity compensation price is high, the remaining carbon quota is loose, and the control objectives tend to be economical and profit maximization;

[0044] In the S2 scenario, both the capacity compensation price and the carbon quota surplus are medium, and the control target tends to be comprehensive balance optimization;

[0045] In the S3 scenario, the capacity compensation price is low, the remaining carbon quota is tight, and the control target tends to prioritize carbon emission reduction.

[0046] Furthermore, it is preferred to encode the optimal load dispatch sequence as the particle position, define a multi-objective fitness function to balance carbon emissions, electricity price costs and capacity benefits, and deal with dispatch constraints by introducing a penalty function:

[0047]

[0048] F(i) (P) = f (i) (P)+Ψ(P)

[0049] Where Ψ(P) represents the particle scheme used to penalize particles that do not meet the constraints; C quota is the total amount of carbon quota; Δt is the duration of a single scheduling cycle; λ1, λ2 are penalty coefficients, is the maximum allowable operating load; Indicates the maximum value of carbon emissions and 0. Indicates the maximum value of energy consumption and 0, f (i) (P) is scene S i The objective function of the control strategy under (i) (P) represents scene S i The complete fitness function under .

[0050] The quantum particle swarm optimization algorithm encodes the load sequence as the particle position and uses the quantum particle swarm optimization algorithm position update formula:

[0051]

[0052] Where, is the position of the i-th particle at the k+1-th iteration time τ, which represents the load value; is the dispatch load value of the i-th particle at time τ; represents the mean of the historical optimal positions of all particles at time τ reflecting the collective search trend, and N is the total number of historical optimal positions of particles; is a random number; β is the quantum scaling factor that controls the search range; is the individual historical optimal position of the i-th particle.

[0053] By running the quantum particle swarm optimization algorithm in multiple scenarios, we can obtain a set of optimal control strategies suitable for different carbon economy scenarios:

[0054] P *(i) =argminF (i) (P),i∈{1,2,3}

[0055] Where, P *(i) For scene S i The optimal control solution obtained by searching.

[0056] Compared with existing technologies, the present invention offers the following advantages: It uses STFormer to accurately predict the carbon emission factor per kilowatt-hour, identifying opportunities for low-carbon regulation. It fully considers capacity compensation and carbon quota mechanisms, combining them to construct a regulation model that balances production stability and carbon cost control. It also utilizes quantum particle swarm optimization to optimize multi-objective control strategies, establishing an integrated low-carbon electricity system for glass factories with high-precision carbon factor prediction capabilities, strong constraints, and optimal regulation. This improves the real-time and overall performance of load low-carbon regulation, facilitating the green transformation of glass factories. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 This is a flow chart of a low-carbon load control method for a glass factory that takes into account capacity compensation and carbon quota, as proposed by the present invention. DETAILED DESCRIPTION

[0058] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0059] A low-carbon load control method for glass plants considering capacity compensation and carbon quota, such as Figure 1 As shown, the method mainly includes the following steps:

[0060] Step 1: Propose a prediction model for the carbon emission factor per kilowatt-hour of glass factories based on the STFormer architecture: This step constructs a multi-dimensional input feature vector that integrates the load characteristics, meteorological environment, electricity price and time information of the glass factory, and uses the STFormer (time series transformer) architecture to predict the carbon emission factor per kilowatt-hour of electricity at multiple moments in the future. The model input includes historical load power, load change rate, ambient temperature, humidity, wind speed, electricity price and clock information. After position encoding to enhance the time series modeling capability, it is sent to the STFormer encoder to extract the time series dependency, and the carbon emission factor per kilowatt-hour of electricity in the future window is output through the decoder. Finally, the carbon emission factor prediction result per kilowatt-hour of electricity at each moment is obtained through regression mapping, providing basic data support for the next step of low-carbon regulation.

[0061] Step 2: A low-carbon load regulation model for the glass factory, taking into account capacity compensation and carbon quotas, is proposed. Based on the carbon emission factor per kilowatt-hour, combined with parameters such as electricity price information, capacity compensation prices, and equipment operating boundaries, a multi-objective low-carbon regulation optimization model is constructed. The model optimizes carbon emissions, minimizes electricity costs, and maximizes capacity compensation benefits, while also considering constraints such as total carbon quota limits, upper and lower load limits, maximum adjustment range, and minimum power guarantees. The optimization variables are the operating loads of the glass factory at each moment within the future scheduling window, and the output is the optimal load scheduling sequence that satisfies the constraints. This model achieves preliminary optimization of flexible scheduling and low-carbon operation based on carbon factor predictions.

[0062] Step three, a low-carbon control strategy for glass factory load based on quantum particle swarm optimization algorithm (QPSO) is proposed: for different operating scenarios (such as high or low capacity compensation price, tight or loose carbon quota, etc.), the quantum particle swarm optimization algorithm is introduced. Based on the optimized load scheme in step two, a dynamic control strategy optimization model is constructed under multiple scenarios. The optimal load scheduling sequence is encoded as the particle position, and a multi-objective fitness function is defined to weigh carbon emissions, electricity price costs and capacity benefits. The scheduling constraints are handled by introducing a penalty function. The quantum particle swarm optimization algorithm is run separately in multiple scenarios to obtain a set of optimal control strategies suitable for different carbon economic scenarios, so as to realize flexible regulation and implementation of low-carbon operation of glass factories.

[0063] For step 1, a multi-dimensional input feature vector is constructed that integrates the glass factory load characteristics, meteorological environment, electricity price, and time information. The STFormer architecture is used to predict the carbon emission factor per kilowatt-hour at multiple future moments. The following steps are included:

[0064] 1) Constructing glass factory load and environmental characteristics

[0065] This step constructs a multidimensional input feature vector for each time point, including the glass factory's load power, load change rate, ambient temperature, humidity, wind speed, electricity price, and time characteristics, comprehensively reflecting the key variables affecting carbon emission factors and providing a basic data structure for subsequent model input.

[0066] Construct a multi-dimensional input feature vector and a historical feature sequence input sample as follows:

[0067]

[0068] X t-T+1:t =[x t-T+1 ,x t-T+2 ,...,x t ] (2)

[0069] Where x t 、x t-T+2 and x t-T+1Represents the input feature vector at different times (including multi-dimensional load characteristics and environmental impact variables), P t is the active power of the glass factory load at time t (kW); is the load change rate at time t; T t is the ambient temperature at time t (℃); H t is the humidity at time t (%); W t is the wind speed at time t (m / s); is the electricity price at time t (yuan / kWh); hour is the current hour (0-23); t day is the current day of the week (Monday to Sunday); X t-T+1:t Represents the historical feature sequence input sample.

[0070] 2) Spatial-temporal position coding

[0071] In order to enhance the STFormer architecture's ability to perceive time sequence, a sine-cosine position encoding mechanism is introduced to inject time series information into the input features, ensuring that the model can still accurately capture the impact of load change trends and environmental disturbances on carbon factors when processing long sequences.

[0072] The multi-dimensional input feature vector is injected into the historical feature sequence input sample, and position encoding is introduced, as shown in the following formula:

[0073]

[0074] The historical feature sequence is input into the sample input position encoding, as shown in the following formula:

[0075]

[0076] In the above formula, e τ is the position encoding vector at time τ, which is used to encode the time sequence information into the model, d is the embedding dimension, i∈[1,d / 2], is the position-encoded input sequence, E t-T+1:t is the position coding sequence from time t-T+1 to time t, that is, E t-T+1:t =[e t-T+1 ,e t-T+2 ,...,e t ].

[0077] 3) STFormer encoding module: spatial-temporal attention modeling

[0078] This part uses the STFormer encoder to capture the long-term temporal dependency between load and environmental variables through the ProbSparse multi-head attention mechanism (sparse attention mechanism), extracts high-dimensional temporal feature representation, and provides rich contextual information representation for the decoding stage.

[0079] Using the sparse attention mechanism, the dependencies between important features in the input sequence are calculated and the input sequence is encoded:

[0080]

[0081] in,

[0082] W Q , W K , W V is the corresponding weight matrix, d k It is the dimension of each vector of Q and K after transformation (i.e. the dimension of query and key), which is used for subsequent calculation of similarity and scaling factor. It plays a normalizing role in attention scoring. Therefore, the constructed input and position encoding directly determine the source of Q, K, V and its dimension d k The superscript T represents the transpose of the matrix. Attention is a calculation function (computation module) used to extract the dependency relationship between time series variables in the input sequence. H enc The high-level representation for each moment is a high-dimensional semantic representation sequence extracted by combining multiple attention layers (usually including multi-head attention, multi-layer stacking, feedforward networks, residual connections, etc.); Enter the sequence after encoding the position. STEnc (·) is the STFormer sequence encoding function, which includes the encoder structure of multi-head attention, feedforward network, and residual connection.

[0083] 4) STFormer decoding module: predicting future carbon emission factors per kilowatt-hour

[0084] By inputting the feature representation extracted by the encoder into the decoder, the carbon emission factor sequence per kilowatt-hour in the future time window is predicted. Each moment of the output corresponds to an estimated carbon factor value, which characterizes the carbon emission level of the glass factory under future operating conditions.

[0085] Generate future forecasts using the historically encoded state:

[0086]

[0087] Where, is the predicted carbon emission factor per kilowatt-hour of electricity in the future (unit: kgCO2 / kWh), H is the future predicted scheduling time (number of moments); F STDec (·) is the STFormer decoder.

[0088] In step 2, based on the carbon emission factor per kilowatt-hour, combined with electricity price information, capacity compensation price, and equipment operation boundary parameters, a multi-objective low-carbon regulation optimization model is constructed:

[0089]

[0090] Where τ is a time point in the prediction window, t is the current time; H is the future prediction scheduling duration (number of time moments); P τ is the optimal dispatch load of the glass factory at time τ (kW), is the predicted carbon emission factor per kilowatt-hour at time τ (kgCO2 / kWh), from step 1; is the electricity price at time τ (yuan / kWh); is the capacity compensation price at time τ (yuan / kW); is the load reduction amplitude, which represents the capacity gain obtained from load reduction, where is the original planned operating load of the glass factory at time τ (kW); α, β, γ are the weight parameters for controlling carbon emissions, electricity costs, and capacity benefits, respectively.

[0091] The multi-objective low-carbon control optimization model is constrained by the total carbon quota constraint, the upper and lower limit constraints of load scheduling, the load adjustment range constraint, and the minimum production line energy consumption constraint of the glass factory. The constraint formula is as follows:

[0092]

[0093] In the above formula, Δt is the duration of a single scheduling cycle (hours), C quota is the total amount of carbon emission quota allowed in the current stage (kgCO2), They are the minimum load boundary and maximum load boundary allowed by the equipment (kW); R max is the maximum load change rate per hour (kW / h); P τ-1 is the dispatching load at the previous moment (kW); E min The minimum total electricity (kWh) required to maintain basic production capacity for the glass factory.

[0094] In step three, the quantum particle swarm optimization algorithm (QPSO) is introduced for different operating scenarios (such as high or low capacity compensation prices, tight or loose carbon quotas, etc.). Based on the optimized load plan in step two, a dynamic control strategy optimization model for multiple scenarios is constructed.

[0095] 1) Control scene classification

[0096] Set up three typical scenarios S i To cope with different actual operating constraints:

[0097] In the S1 scenario, the capacity compensation price is high, the remaining carbon quota is loose, and the control objectives tend to be economical and profit maximization;

[0098] In the S2 scenario, both the capacity compensation price and the carbon quota surplus are medium, and the control target tends to be comprehensive balance optimization;

[0099] In the S3 scenario, the capacity compensation price is low, the remaining carbon quota is tight, and the control target tends to prioritize carbon emission reduction;

[0100] Among them, C cap is the capacity compensation price, C quota is the remaining carbon quota, and the weight combination (α i ,β i ,γ i ), indicating that in scene S i The weights of the three types of targets (carbon, electricity charges, and capacity compensation) under the above model are as follows: i = 1, 2, and 3.

[0101] 2) Particle encoding and initialization

[0102] Introducing quantum particle swarm optimization algorithm, particle position The scheduling of the i-th particle controls the load in the future H period, which represents the control strategy of the glass factory within the prediction window, that is, the load scheduling sequence. When initializing the population, a disturbance is added, as shown in the following formula:

[0103]

[0104] Where, is the dispatch load value of the i-th particle at time τ; The optimized scheduling value obtained by the multi-objective low-carbon regulation optimization model is used as a reference for particle initialization; is the initial disturbance term, satisfying ε~U(-δ,δ) to ensure population diversity, U(-δ,δ) is the disturbance range, and δ is the disturbance threshold.

[0105] 3) Multi-objective fitness function design

[0106] The dynamic control strategy optimization model constructed in multiple scenarios is as follows:

[0107]

[0108] Where τ is a time point within the prediction window; t is the current time; H is the future prediction scheduling duration (number of moments); P τ is the optimal dispatch load of the glass factory at time τ (kW); To predict the carbon emission factor per kilowatt-hour, the first step is to predict the result; is the electricity price at time τ (yuan / kWh); is the capacity compensation price at time τ (yuan / kW); is the load reduction amplitude, which represents the capacity gain obtained from load reduction, where is the original planned operating load of the glass factory at time τ; (α i ,β i ,γ i ) is a weight combination, indicating that in three typical scenarios S i The weights under, where i = 1, 2, 3.

[0109] 4) Constraint handling method

[0110] The optimal load dispatch sequence is encoded as the particle position, and a multi-objective fitness function is defined to balance carbon emissions, electricity price costs, and capacity benefits. The dispatch constraints are handled by introducing a penalty function. The penalty function method is used to incorporate the constraints in step (2) into the fitness function. The penalty term is constructed as shown in formula (15), and the final fitness is shown in formula (16):

[0111]

[0112] F (i) (P) = f (i) (P)+Ψ(P) (16)

[0113] Where Ψ(P) represents the particle scheme used to penalize particles that do not meet the constraints; C quota is the total amount of carbon quota (kgCO2); Δt is the duration of a single scheduling cycle (hours); λ1 and λ2 are penalty coefficients that adjust the impact of constraint violation on fitness; is the maximum allowable operating load (kW); Indicates the maximum value of carbon emissions and 0. Indicates the maximum value of energy consumption and 0, f (i) (P) is scene S i The objective function of the control strategy under (to be minimized), F (i) (P) represents scene S i The complete fitness function under .

[0114] 5) Quantum Particle Swarm Optimization (QPSO) Position Update Mechanism (Cross-Scenario Control)

[0115] The quantum particle swarm optimization algorithm encodes the load sequence as the particle position and uses the quantum particle swarm optimization algorithm position update formula:

[0116]

[0117] Where, P τ (i)(k+1) is the position of the i-th particle at the k+1-th iteration time τ, indicating the load value; is the dispatch load value of the i-th particle at time τ; represents the mean of the historical optimal positions of all particles at time τ reflecting the collective search trend, and N is the total number of historical optimal positions of particles; is a random number, which introduces uncertainty and enhances jumpiness; β is the quantum scaling factor that controls the search range; is the individual historical optimal position (individual extreme value) of the i-th particle.

[0118] 6) Multi-scenario optimization output control strategy set

[0119] By running the quantum particle swarm optimization algorithm in multiple scenarios, we can obtain a set of optimal control strategies suitable for different carbon economy scenarios:

[0120] P *(i) =argminF (i) (P), i∈{1,2,3} (18)

[0121] Where, P *(i) For scene S i The optimal control scheme obtained by searching under *(1) 、P *(2) 、P *(3) They are the control strategy under high capacity benefit scenario (economic type), the compromise control strategy under medium conditions (balanced type), and the control strategy under carbon constraint scenario (low carbon type); F (i) (P) is scene S i The complete fitness function under , including target and penalty.

[0122] It should be noted that the parts not covered by the present invention are the same as the existing technology or can be implemented by using the existing technology. The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art who, within the technical scope disclosed by the present invention, makes equivalent substitutions or changes based on the technical solution and inventive concept of the present invention shall be covered by the scope of protection of the present invention.

Claims

1. A low-carbon control method for glass factory load considering capacity compensation and carbon quota, characterized in that: The following steps are involved: Step 1: Construct a multi-dimensional input feature vector that integrates the glass factory's load characteristics, meteorological environment, electricity price, and time information. Use the STFormer architecture to predict the carbon emission factors per kilowatt-hour at multiple future moments, extract the temporal dependencies, and output the carbon emission factors per kilowatt-hour within the future window through a decoder. Step 2: Based on the carbon emission factor per kilowatt-hour, combined with electricity price information, capacity compensation price, and equipment operation boundary parameters, a multi-objective low-carbon regulation optimization model is constructed; Step three: For different operating scenarios, the quantum particle swarm optimization algorithm is introduced to construct a dynamic control strategy optimization model under multiple scenarios. The optimal load scheduling sequence is encoded as the particle position, and a multi-objective fitness function is defined to balance carbon emissions, electricity price costs and capacity benefits. The scheduling constraints are handled by introducing a penalty function. The quantum particle swarm optimization algorithm is run separately in multiple scenarios to obtain a set of optimal control strategies suitable for different carbon economic scenarios.

2. The low-carbon control method for glass factory load considering capacity compensation and carbon quota according to claim 1 is characterized in that: In step 1, a multi-dimensional input feature vector is constructed that integrates the glass factory load characteristics, meteorological environment, electricity price and time information. The following steps are involved: Construct a multi-dimensional input feature vector and a historical feature sequence input sample as follows: X t-T+1:t =[x t-T+1 ,x t-T+2 ,...,x t ] Where x t 、x t-T+2 and x t-T+1 Represents the input feature vector at different times, P t is the active power of the glass factory load at time t; is the load change rate at time t; T t is the ambient temperature at time t; H t is the humidity at time t; W t is the wind speed at time t; is the electricity price at time t; hour is the current hour; t day is the current day of the week; X t-T+1:t Represents the historical feature sequence input sample.

3. The low-carbon control method for glass factory load considering capacity compensation and carbon quota according to claim 2 is characterized in that: The multi-dimensional input feature vector is injected into the historical feature sequence input sample, and position encoding is introduced, as shown in the following formula: The historical feature sequence is input into the sample input position encoding, as shown in the following formula: In the above formula, e τ is the position encoding vector at time τ, d is the embedding dimension, i∈[1,d / 2], is the position-encoded input sequence, E t-T+1:t is the position code sequence from time t-T+1 to time t, E t-T+1:t =[e t-T+1 ,e t-T+2 ,...,e t ]; Using the sparse attention mechanism, the dependencies between important features in the input sequence are calculated and the input sequence is encoded: in, W Q , W K , W V is the corresponding weight matrix, d k is each vector dimension of Q and K after conversion. Attention is a calculation function for extracting the temporal variable dependency between each moment in the input sequence. H enc Represents the higher order of each moment; is the position-encoded input sequence, F STEnc (·) is the STFormer sequence encoding function, and the superscript T represents the transpose of the matrix; By inputting the feature representation extracted by the encoder into the decoder, the carbon emission factor sequence of each kilowatt-hour in the future time window is predicted. Each moment of the output corresponds to an estimated carbon factor value, which describes the carbon emission level of the glass factory under the future operating state. Generate future forecasts using the historically encoded state: Where, is the carbon emission factor per kilowatt-hour of electricity predicted at a future time, H is the future predicted scheduling duration; F STDec (·) is the STFormer decoder.

4. The low-carbon control method for glass factory load considering capacity compensation and carbon quota according to claim 1 is characterized in that: In step 2, based on the carbon emission factor per kilowatt-hour, combined with electricity price information, capacity compensation price, and equipment operation boundary parameters, a multi-objective low-carbon regulation optimization model is constructed: Where τ is a time point in the prediction window, t is the current time; H is the future prediction scheduling duration; P τ is the optimal dispatch load of the glass factory at time τ, is the predicted carbon emission factor per kilowatt-hour at time τ; is the electricity price at time τ; is the capacity compensation price at time τ; To reduce the load range, is the original planned operating load of the glass factory at time τ; α, β, γ are the weight parameters for controlling carbon emissions, electricity costs, and capacity benefits, respectively.

5. The low-carbon control method for glass factory load considering capacity compensation and carbon quota according to claim 4 is characterized in that: The multi-objective low-carbon control optimization model is constrained by the total carbon quota constraint, the upper and lower limit constraints of load scheduling, the load adjustment range constraint, and the minimum production line energy consumption constraint of the glass factory. The constraint formula is as follows: In the above formula, Δt is the duration of a single scheduling cycle, C quota is the total amount of carbon emission quota allowed in the current stage, are the minimum load boundary and maximum load boundary allowed by the equipment respectively; R max P is the maximum load change rate per hour; τ-1 is the dispatching load at the previous moment; E min The minimum total electricity required to maintain basic production capacity for a glass factory.

6. The low-carbon control method for glass factory load considering capacity compensation and carbon quota according to claim 1 is characterized in that: In step three, the quantum particle swarm optimization algorithm is introduced to construct a dynamic control strategy optimization model for different operating scenarios: Introducing quantum particle swarm optimization algorithm, particle position The scheduling control of the i-th particle controls the load in the future H period, which represents the control strategy of the glass factory within the prediction window. When initializing the population, disturbance is added, as shown in the following formula: Where, is the dispatch load value of the i-th particle at time τ; The optimized dispatch value obtained by the multi-objective low-carbon regulation optimization model; is the initial disturbance term, satisfying ε~U(-δ,δ), U(-δ,δ) is the disturbance range, and δ is the disturbance threshold; The dynamic control strategy optimization model constructed in multiple scenarios is as follows: Where τ is a time point within the prediction window; t is the current time; H is the future prediction scheduling duration; P τ is the optimal dispatch load of the glass factory at time τ; To predict the carbon emission factor per kilowatt-hour; is the electricity price at time τ; is the capacity compensation price at time τ; is the load reduction amplitude, which represents the capacity gain obtained from load reduction, where is the original planned operating load of the glass factory at time τ; (α i ,β i ,γ i ) is a weight combination, indicating that in three typical scenarios S i The weights under, where i = 1, 2, 3.

7. The low-carbon control method for glass factory load considering capacity compensation and carbon quota according to claim 6 is characterized in that: According to different operating scenarios, three typical scenarios S are set i To cope with different actual operating constraints: In the S1 scenario, the capacity compensation price is high, the remaining carbon quota is loose, and the control objectives tend to be economical and profit maximization; In the S2 scenario, both the capacity compensation price and the carbon quota surplus are medium, and the control target tends to be comprehensive balance optimization; In the S3 scenario, the capacity compensation price is low, the remaining carbon quota is tight, and the control target tends to prioritize carbon emission reduction.

8. The low-carbon control method for glass factory load considering capacity compensation and carbon quota according to claim 6 is characterized in that: The optimal load dispatch sequence is encoded as the particle position, and a multi-objective fitness function is defined to balance carbon emissions, electricity price costs, and capacity benefits. The dispatch constraints are handled by introducing a penalty function: F (i) (P)=f (i) (P)+Ψ(P) Where Ψ(P) represents the particle scheme used to penalize particles that do not meet the constraints; C quota is the total amount of carbon quota; Δt is the duration of a single scheduling cycle; λ1, λ2 are penalty coefficients, is the maximum allowable operating load; Indicates the maximum value of carbon emissions and 0. Indicates the maximum value of energy consumption and 0, f (i) (P) is scene S i The objective function of the control strategy under (i) (P) represents scene S i The complete fitness function under ; The quantum particle swarm optimization algorithm encodes the load sequence as the particle position and uses the quantum particle swarm optimization algorithm position update formula: Where, is the position of the i-th particle at the k+1-th iteration time τ, which represents the load value; is the dispatch load value of the i-th particle at time τ; represents the mean of the historical optimal positions of all particles at time τ reflecting the collective search trend, and N is the total number of historical optimal positions of particles; is a random number; β is the quantum scaling factor that controls the search range; P τ (i) (k) is the individual historical optimal position of the i-th particle; By running the quantum particle swarm optimization algorithm in multiple scenarios, we can obtain a set of optimal control strategies suitable for different carbon economy scenarios: P *(i) =argminF (i) (P),i∈{1,2,3} Where, P *(i) For scene S i The optimal control solution obtained by searching.

Images