Industrial air conditioner load scheduling method based on improved market supervision optimization algorithm

Abstract

The invention belongs to the technical field of industrial air conditioner load scheduling, and particularly relates to an industrial air conditioner load scheduling method based on an improved market supervision optimization algorithm, which comprises the following steps: S1, collecting historical meteorological data and historical photovoltaic power generation output data, and inputting the data into a Transform-GRU hybrid prediction model to obtain a prediction sequence of future preset duration; s2, constructing an air conditioner load scheduling optimization model based on the prediction sequence; s3, solving by adopting an improved market supervision optimization algorithm to obtain an air conditioner load scheduling strategy and a corresponding dynamic temperature and humidity set value; s4, the air conditioner load dispatching strategy and the dynamic temperature and humidity set value are issued to environment regulation and control equipment to be executed; and S5, collecting current actual meteorological data and photovoltaic power generation output data at a preset time interval, taking the current actual meteorological data and the photovoltaic power generation output data as the latest historical data, updating the latest historical data to the input in the step S1, and executing the steps from the step S1 to the step S4 again. According to the method, organic unification of three targets of efficient consumption of photovoltaic power generation, remarkable reduction of the operation cost and stable guarantee of the process environment can be realized.

Description

Technical Field

[0001] This invention belongs to the field of industrial air conditioning load scheduling technology, and particularly relates to an industrial air conditioning load scheduling method based on an improved market supervision optimization algorithm. Background Technology

[0002] Energy management in the industrial sector, particularly the energy consumption control of air conditioning systems in high-energy-consuming enterprises such as cigarette factories, has always been a key focus and challenge in achieving energy conservation and emission reduction under the "dual carbon" target. Cigarette production has stringent process requirements for workshop temperature and humidity environments, and air conditioning systems operate at high loads year-round, consuming over 30% of the factory's total electricity. Meanwhile, with the rapid adoption of distributed photovoltaic and other renewable energy sources in industrial settings, how to efficiently absorb the highly volatile and intermittent nature of photovoltaic power generation and dynamically coordinate it with the needs of high-precision environmental control has become a crucial issue for the intelligent upgrading of current industrial energy systems.

[0003] Existing technologies mainly focus on three aspects: photovoltaic power generation forecasting, air conditioning load regulation, and optimized scheduling algorithms. However, they still have significant limitations in practical applications, making it difficult to meet the comprehensive requirements of economy, stability, and flexible regulation capabilities in complex industrial scenarios. The specific problems can be summarized into the following three points:

[0004] (1) Insufficient specificity and limited accuracy of prediction models. Current mainstream photovoltaic power generation and meteorological parameter prediction methods mostly adopt general time series models (such as LSTM, standard Transformer, etc.), which do not fully consider the strong coupling relationship between local micro-meteorological conditions (such as factory building shading, heat island effect) and photovoltaic output in specific industrial scenarios such as cigarette factories. When dealing with multivariable, non-stationary, and high-noise industrial measured data, such models often have problems such as insufficient feature extraction and weak long-term reliance on modeling capabilities, resulting in high prediction errors (such as RMSE and MAPE indices being significantly worse than industry requirements). Prediction deviation directly affects the reliability of subsequent scheduling decisions, causing the optimization results to deviate from the actual operating conditions, and even causing control failure.

[0005] (2) Poor convergence performance of optimization algorithms and neglect of multiple constraints. Traditional intelligent optimization algorithms (such as particle swarm optimization, genetic algorithms, or standard market supervision optimization algorithms) generally suffer from defects such as rapid population diversity decay and easy getting trapped in local optima when solving high-dimensional, non-convex, and multi-constrained air conditioning load scheduling problems. Especially when facing multiple coupled constraints such as time-of-use pricing, equipment start-up and shutdown characteristics, and hard process boundaries of temperature and humidity, the algorithms have difficulty effectively balancing exploration and development capabilities, resulting in low-quality or infeasible solution sets. More importantly, most existing scheduling models treat temperature and humidity control only as soft constraints or ex-post verification items, without embedding it into the core constraint system of the optimization framework, causing some "economically optimal" solutions to violate production process specifications in actual implementation and pose safety hazards.

[0006] (3) Rigid control strategy, lacking flexible adjustment capability. Current industrial air conditioning systems generally adopt control logic with fixed temperature and humidity setpoints, lacking the ability to dynamically adjust the load according to the supply of renewable energy. This rigid strategy cannot take advantage of the peak output of photovoltaic power to actively "absorb" excess green electricity (for example, by appropriately relaxing the temperature and humidity settings to increase the cooling / dehumidification load), nor can it pre-cool / preheat in advance during the photovoltaic off-peak period to reduce the purchase of electricity from the grid at high prices. As a result, the self-consumption rate of photovoltaic power is low, the system is highly dependent on the grid, the overall energy economy is limited, and it is difficult to achieve flexible interaction between source and load.

[0007] Therefore, how to achieve the organic unity of the three goals of efficient photovoltaic power generation, significant reduction in operating costs, and stable process environment has become an urgent problem to be solved. Summary of the Invention

[0008] To address the shortcomings of the existing technologies, this invention provides an industrial air conditioning load scheduling method based on an improved market supervision optimization algorithm, which can achieve the organic unity of the three objectives of efficient photovoltaic power generation, significant reduction in operating costs, and stable process environment.

[0009] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0010] An industrial air conditioning load scheduling method based on an improved market supervision optimization algorithm includes the following steps:

[0011] S1. Collect historical meteorological data and historical photovoltaic power generation output data, input them into a pre-trained Transformer-GRU hybrid prediction model, and obtain a prediction sequence of photovoltaic power generation and outdoor environmental parameters for a preset duration in the future; wherein, the Transformer-GRU hybrid prediction model embeds a gated recurrent unit (GRU) layer in the standard Transformer encoder-decoder architecture, and enhances the temporal feature extraction capability through an additive fusion mechanism.

[0012] S2. Based on the predicted sequence obtained in S1, an air conditioning load scheduling optimization model is constructed. The objective function of the model includes two terms: net load cost and temperature / humidity over-limit penalty function, with the goal of minimizing the sum of the objective functions. The net load cost is calculated based on the time-of-use electricity price and the power grid interaction, where the power grid interaction is the power consumption of the air conditioning system minus the photovoltaic power generation. The temperature / humidity over-limit penalty function is used to penalize the cumulative number of times the actual indoor temperature and humidity deviate from the set range. The optimization model uses the indoor temperature and humidity set value range allowed by the process environment as a hard constraint and uses the predicted value of photovoltaic power generation as a real-time available power supply capacity constraint.

[0013] S3. The improved market supervision optimization algorithm is used to solve the optimization model constructed in S2 to obtain the air conditioning load scheduling strategy and the corresponding dynamic temperature and humidity setpoints.

[0014] The improved market supervision optimization algorithm constructs a collaborative optimization framework encompassing four key roles: simulated business management, market transactions, fairness mechanisms, and feedback mechanisms. Enhancements are achieved within this framework through the following methods:

[0015] a) During the population exploration phase, the management role, market transaction role, fairness mechanism role and feedback mechanism role respectively execute differentiated search strategies, and adaptive random perturbation is introduced during the individual update process to enhance the global exploration capability;

[0016] b) During the local development phase, perform gradient-guided crossover and mutation operations on dominant individuals to improve convergence accuracy;

[0017] c) During the algorithm iteration process, the population diversity index is monitored in real time. When the index is lower than the preset threshold, the population dynamic recombination mechanism is triggered, and the feedback mechanism is used to guide the population out of the local optimum to prevent premature convergence.

[0018] S4. Send the air conditioning load scheduling strategy and dynamic temperature and humidity setpoints obtained in S3 to the environmental control equipment for execution. By dynamically adjusting the temperature and humidity setpoints within the allowable range of the process, actively shape the air conditioning load curve so that it matches the fluctuation characteristics of photovoltaic power generation in the time dimension.

[0019] S5. Collect current actual meteorological data and photovoltaic power generation output data at preset time intervals, use them as inputs to update the latest historical data in S1, and re-execute steps S1 to S4 to form a dynamic rolling optimization closed loop to continuously ensure the stability of the process environment and energy economy.

[0020] Compared with the prior art, the present invention has the following beneficial effects:

[0021] 1. Enhance the timing matching capability between photovoltaic power generation and air conditioning load. By introducing a dynamic temperature and humidity setting mechanism, the system can proactively adjust the air conditioning load curve while meeting process constraints, enabling it to better track the fluctuation characteristics of photovoltaic power generation over time. Compared to the traditional rigid control strategy with fixed setpoints, this method significantly enhances the load-side response flexibility to photovoltaic output, effectively increases the proportion of photovoltaic self-consumption, and reduces dependence on high-priced grid electricity.

[0022] 2. Enhanced Prediction Accuracy and Scene Adaptability. The Transformer-GRU hybrid prediction model adopted integrates the global attention mechanism of Transformer with the local temporal modeling capability of GRU, and strengthens key feature extraction through an additive fusion mechanism. Compared with general prediction models with single structures (such as standard LSTM or pure Transformer), this model is more suitable for the complex data characteristics of strong coupling between photovoltaic power output and micro-meteorology in industrial scenarios such as cigarette factories, thus providing a more reliable input basis for subsequent optimization.

[0023] 3. Improve the stability and global search capability of the optimization solution. The improved market supervision optimization algorithm constructs a collaborative framework of four roles, introduces differentiated search strategies and adaptive perturbation mechanisms in the population exploration phase, performs refined operations by combining gradient-guided information in the local development phase, and is supplemented by population diversity monitoring and dynamic recombination mechanisms. Compared with traditional particle swarm optimization, genetic algorithms, or standard market supervision algorithms, this method effectively alleviates the premature convergence problem and can still stably converge to a high-quality feasible solution under high-dimensional nonlinearity and multi-constraint conditions.

[0024] 4. Ensure compliance with process environment and feasibility of system operation. Temperature and humidity setting ranges are directly embedded as hard constraints into the optimization model, ensuring that all scheduling strategies strictly meet production process requirements. This avoids the problem of some existing economy-oriented scheduling methods becoming unexecutable due to neglecting hard constraints, significantly improving the engineering practicality and field deployability of the optimization results.

[0025] 5. Achieve closed-loop rolling optimization and continuous adaptive control. By periodically updating actual operating data and re-executing the prediction-optimization-execution process, the system can dynamically adapt to uncertainties such as sudden weather changes, equipment aging, or load disturbances, forming a robust rolling scheduling mechanism. Compared to static or single-time optimization schemes, this closed-loop structure significantly enhances the system's long-term operational stability and energy management intelligence.

[0026] In summary, this method can achieve the organic unity of the three objectives of efficient photovoltaic power generation, significant reduction in operating costs, and stable process environment.

[0027] Preferably, in S2, the decision variables of the optimization model include the dynamically changing power consumption of the air conditioning system, the indoor temperature setpoint, and the indoor relative humidity setpoint within a preset time period in the future.

[0028] This setup differs from traditional methods, which typically rely solely on equipment start-up / shutdown or fixed temperature and humidity settings for scheduling. Instead, this solution incorporates the temperature and humidity setpoints themselves as adjustable variables into the optimization framework, enabling the system to proactively "shape" the load curve within the limits allowed by the technological process. This collaborative modeling approach significantly enhances the scheduling strategy's responsiveness to fluctuations in photovoltaic power generation.

[0029] 2. Enhancing the balance between operational economy and control flexibility. By jointly optimizing power consumption and environmental setpoints, the algorithm can appropriately relax temperature and humidity settings (within process boundaries) during periods of low electricity prices or peak photovoltaic output to increase green electricity consumption or reduce electricity purchase costs; conversely, it can tighten settings during critical periods to ensure environmental quality. Compared to methods that only optimize power consumption or only adjust equipment status, this strategy more fully taps into the potential of flexible control.

[0030] Preferably, the relationship between the power consumption of the air conditioning system and the outdoor temperature and humidity and the indoor temperature and humidity setpoints is determined by the following quantitative model of air conditioning load:

[0031] ;

[0032] In the formula, Let be the power consumption of the air conditioning system at time t; and These represent the outdoor temperature and relative humidity at time t, as predicted by S1. and These are the indoor temperature setting and indoor relative humidity setting at time t, respectively.

[0033] This setup achieves two key advantages: 1. Improved accuracy and predictability of air conditioning load modeling. The model explicitly expresses air conditioning power consumption as a nonlinear function of outdoor temperature and humidity versus indoor setpoints, fully characterizing the impact of environmental parameter changes on energy consumption. Compared to empirical estimation or black-box simulation methods, this analytical model offers higher transparency and physical consistency, providing a reliable power consumption calculation basis for optimized scheduling.

[0034] 2. Supports real-time load response analysis under dynamic temperature and humidity settings. The model explicitly includes indoor temperature and humidity setpoints as variables, allowing direct evaluation of the impact of different setting strategies on system power consumption during optimization. This supports the flexible control logic of "controlling load by adjusting temperature," enhancing the scientific nature and operability of scheduling decisions.

[0035] This quantitative model of air conditioning load provides a key mathematical tool for achieving source-load coordinated optimization, effectively improving the system's energy management accuracy and response capability under complex operating conditions.

[0036] Preferably, in each iteration of S3, the search weight coefficients configured for the management role, market transaction role, fairness mechanism role, and feedback mechanism role are dynamically and adaptively adjusted based on the current iteration number and the rate of change of the population's historical best solution, in order to balance global exploration and local development capabilities.

[0037] This setup has two advantages: 1. It enhances the algorithm's adaptability at different optimization stages. The search weight coefficients are adjusted in real time based on the current iteration progress and the rate of change of the population's historical best solutions. This allows the algorithm to focus more on global exploration in the early stages (e.g., dominating management and market transaction roles) and more on local fine-tuning in the later stages (e.g., strengthening the roles of fairness and feedback mechanisms). Compared to strategies with fixed weights or those that rely solely on the number of iterations, this mechanism can respond more sensitively to the actual state of the optimization process, improving search efficiency.

[0038] 2. Effectively reconcile the contradiction between diversity and convergence. By introducing the rate of change of historical optimal solutions as a feedback signal, the weight of the exploratory role is automatically increased when optimization stagnates, avoiding premature convergence; while during the rapid performance improvement phase, the development capability is strengthened to accelerate the convergence speed. This dynamic balancing mechanism is significantly superior to traditional static or heuristic weight allocation methods, improving the overall robustness and solution quality of the algorithm.

[0039] This adaptive weight adjustment mechanism enables the improved market supervision optimization algorithm to have both greater exploration breadth and development depth in complex industrial scheduling scenarios, providing key support for high-precision and high-reliability load scheduling.

[0040] Preferably, in step S3a), the random perturbation is achieved by superimposing noise following a Gaussian distribution onto the individual position vector, wherein the standard deviation of the Gaussian distribution adaptively decreases with the increase of the number of iterations.

[0041] This setup achieves two key advantages: 1) It balances early-stage exploration capability with later-stage convergence stability. In the early stages of the algorithm, a larger standard deviation of Gaussian noise enhances the intensity of random perturbations in the population, helping to escape localized regions and expand the search range. As iterations progress, the standard deviation gradually decays, reducing the amplitude of perturbations and allowing individuals to more stably cluster towards regions with high-quality solutions. Compared to random perturbations of fixed intensity, this strategy more effectively matches the needs of different stages in the optimization process.

[0042] 2. Improve algorithm robustness and suppress premature convergence. The adaptive decay mechanism avoids oscillations or convergence failures caused by excessive perturbations in the later stages, and also prevents loss of diversity due to insufficient perturbations. Compared with traditional constant or linear perturbation methods, this method maintains population vitality while ensuring convergence accuracy, effectively alleviating premature convergence in complex multi-constraint scheduling problems.

[0043] Preferably, in step S3b), the gradient guidance information is obtained by performing a finite difference approximation of the objective function in the neighborhood of the current dominant individual, and is used to guide the direction and step size of crossover and mutation operations.

[0044] This setup improves the directionality and efficiency of local search. Traditional evolutionary algorithms typically rely on random operations for crossover and mutation, lacking awareness of the local shape of the objective function. This scheme estimates the local gradient direction by performing finite difference approximation within the neighborhood of dominant individuals, directing subsequent operations in a direction that is more likely to reduce operating costs, significantly enhancing the targeting and convergence speed of local exploration.

[0045] 2. Achieving gradient-like optimization without relying on explicit gradients. Industrial air conditioning scheduling models often contain non-differentiable or black-box components (such as equipment start-up and shutdown logic, discrete control variables), making it difficult to directly apply gradient-based optimization methods. This scheme implicitly obtains gradient information through finite difference, retaining the global framework of evolutionary algorithms while integrating the fine-tuning capabilities of gradient-like methods, thus balancing flexibility and accuracy.

[0046] This mechanism effectively bridges the gap between gradient-free intelligent optimization and gradient-driven local search, enabling the improved market supervision optimization algorithm to possess both strong robustness and high solution accuracy in complex, non-smooth industrial scheduling problems.

[0047] Preferably, in step S3c), the population diversity index is calculated using the average Euclidean distance between all individuals in the population, and the preset threshold is set to 10% to 50% of the population diversity index value when the algorithm is initialized.

[0048] This setup achieves two key advantages: 1. Sensitive monitoring and objective assessment of population aggregation status. Mean Euclidean distance effectively reflects the breadth of individual distribution in the solution space, providing a more accurate depiction of the spatial dispersion of the population compared to diversity indicators that rely solely on fitness variance or simple counting. 2. Combining this with a relative threshold based on initial diversity avoids the problem of poor adaptability of fixed thresholds across different problem sizes, making diversity assessment more robust and transferable.

[0049] 2. Precise triggering of the recombination mechanism effectively avoids premature convergence. Setting the threshold to a proportion of initial diversity (10%–50%) ensures that the algorithm initiates local development only after sufficient exploration, while also activating the feedback mechanism for dynamic recombination when the population aggregates too early. Compared to no threshold control or empirically set absolute thresholds, this strategy better aligns with the actual evolutionary patterns of the optimization process, significantly improving the ability to escape local optima.

[0050] This diversity assessment and triggering mechanism provides a reliable "health diagnosis" method for the improved market supervision optimization algorithm, enhancing its global search stability and solution quality assurance capabilities in high-dimensional, multi-constraint industrial scheduling scenarios.

[0051] Preferably, in step S3c), when the population dynamic recombination mechanism is triggered, the current best elite individuals are retained, and the remaining individuals are randomly initialized or perturbed based on the information of the elite individuals, so as to reconstruct the population and maintain global exploration capability.

[0052] This setup balances the inheritance of high-quality solutions with the restoration of population diversity. Retaining a few currently optimal elite individuals ensures that existing high-quality solutions are not discarded, preventing the loss of optimization results; simultaneously, non-elite individuals are initialized randomly or with perturbations based on elite information, effectively injecting new solution space regions and quickly restoring population diversity. Compared to a full population reset or a completely random restart, this strategy achieves a better balance between exploration and development.

[0053] 2. Enhances the algorithm's ability to escape local optima and improves convergence robustness. When premature convergence is detected in the population, this recombination mechanism can quickly break the clustering of individuals and guide the search process to expand back to the less explored regions. Especially in high-dimensional, multi-peak air conditioning load scheduling problems, this method significantly enhances the algorithm's ability to escape local traps and improves the stability and reliability of the final solution.

[0054] This elite retention and intelligent reconstruction strategy provides an efficient and robust mechanism for dynamic population updates, and is a key element in ensuring the continuous and efficient operation of the improved market supervision optimization algorithm in complex industrial optimization scenarios.

[0055] Preferably, in S1, in the Transformer-GRU hybrid prediction model, the GRU layer is embedded after the input position encoding of the Transformer encoder; the additive fusion mechanism is implemented through an additive fusion layer, which is used to fuse the output features of the Transformer encoder and the intermediate feature representation of the decoder; the Transformer-GRU hybrid prediction model also includes a Dropout regularization layer and a fully connected regression output layer, wherein the deactivation rate of the Dropout regularization layer is set to 0.1.

[0056] This setup enhances the multi-level modeling capability of temporal features. By placing the GRU layer after positional encoding and before the main encoder, the model captures local temporal dependencies before entering self-attention computation, effectively compensating for the lack of sensitivity of pure Transformer to short-term dynamic changes. At the same time, the additive fusion mechanism explicitly fuses the high-level semantics of the encoder with the intermediate state of the decoder, strengthening the temporal consistency between historical information and future predictions, and improving the overall prediction accuracy.

[0057] 2. Improve model generalization ability and suppress overfitting. A lightweight Dropout regularization with an inactivation rate of only 0.1 is introduced to moderately suppress overfitting to the training data while preserving the model's expressive power. This is especially suitable for photovoltaic and meteorological data with limited sample size but complex noise in industrial scenarios, enhancing the robustness and stability of the model in actual operation.

[0058] This structural design organically integrates local loop modeling, global attention mechanism and regularization control, enabling the prediction model to maintain high expressiveness while possessing stronger adaptability and reliability, and providing high-quality input guarantee for subsequent optimization and scheduling.

[0059] Preferably, in S2, the allowable indoor temperature and humidity setting range for the process environment is a temperature of 25°C to 29°C and a relative humidity of 55%RH to 65%RH. Attached Figure Description

[0060] To make the objectives, technical solutions, and advantages of the invention clearer, the invention will now be described in further detail with reference to the accompanying drawings, wherein:

[0061] Figure 1 This is a flowchart of the method;

[0062] Figure 2 This is a schematic diagram illustrating the workflow of each model in this method;

[0063] Figure 3 This is a flowchart illustrating the Transformer-GRU hybrid prediction model in Example 1.

[0064] Figure 4 This is a flowchart of the improved market supervision optimization algorithm in Example 1;

[0065] Figure 5 This is a comparison chart of the prediction results in Example 2 of Example 2. Detailed Implementation

[0066] The following detailed explanation illustrates the specific implementation methods:

[0067] Example 1

[0068] like Figure 1 , Figure 2 As shown, this embodiment discloses an industrial air conditioning load scheduling method based on an improved market supervision optimization algorithm, including the following steps:

[0069] S1. Collect historical meteorological data and historical photovoltaic power generation data, and input them into a pre-trained Transformer-GRU hybrid prediction model to obtain a predicted sequence of photovoltaic power generation and outdoor environmental parameters for a preset future duration. The Transformer-GRU hybrid prediction model embeds a gated recurrent unit (GRU) layer into the standard Transformer encoder-decoder architecture and enhances its temporal feature extraction capability through an additive fusion mechanism. The preset future duration is 24 hours.

[0070] In specific implementation, in the Transformer-GRU hybrid prediction model, the GRU layer is embedded after the input position encoding of the Transformer encoder; the additive fusion mechanism is implemented through an additive fusion layer, which is used to fuse the output features of the Transformer encoder and the intermediate feature representation of the decoder; the Transformer-GRU hybrid prediction model also includes a Dropout regularization layer and a fully connected regression output layer, wherein the deactivation rate of the Dropout regularization layer is set to 0.1.

[0071] By placing the GRU layer after positional encoding and before the main encoder, the model captures local temporal dependencies before entering self-attention computation, effectively compensating for the lack of sensitivity to short-term dynamic changes in pure Transformers. Simultaneously, the additive fusion mechanism explicitly fuses the high-level semantics of the encoder with the intermediate states of the decoder, strengthening the temporal consistency between historical information and future predictions and improving overall prediction accuracy. Furthermore, the introduction of lightweight Dropout regularization with a deactivation rate of only 0.1 moderately suppresses overfitting to training data while preserving the model's expressive power. This is particularly suitable for photovoltaic and meteorological data with limited sample size but complex noise in industrial scenarios, enhancing the model's robustness and stability in real-world operation.

[0072] This method employs a novel Transformer-GRU hybrid neural network architecture, deeply integrating the temporal modeling capabilities of GRU (Gated Recurrent Unit) with the feature association capture capabilities of Transformer's multi-head self-attention mechanism. The model effectively learns long-term and short-term temporal dependencies in meteorological and load data through GRU layers, and then focuses on key features in parallel through the multi-head attention mechanism, significantly improving the prediction accuracy for strong temporal series and non-stationary data such as photovoltaic power output and outdoor temperature and humidity. Its structure sequentially includes an input embedding and position encoding layer (preserving sequence order information), an additive layer (fusing multiple feature paths), a dropout layer (random deactivation to prevent overfitting), and a fully connected regression output layer. The overall design, while ensuring model expressive power, improves computational efficiency and generalization performance through structural simplification and regularization strategies. During prediction, the model uses historical and real-time meteorological data (such as temperature and humidity) as input, and achieves multi-step high-precision predictions of future photovoltaic power generation and outdoor temperature and humidity (output) through end-to-end training, providing reliable input for subsequent optimized scheduling.

[0073] The usage process of the Transformer-GRU hybrid prediction model is as follows: Figure 3 The diagram illustrates how to derive the final forecast from raw meteorological data. It emphasizes the modeling of the correlation between meteorological data and photovoltaic power generation, showcasing the unique approach of this invention to indirectly improve the accuracy of photovoltaic power generation forecasts using meteorological trend predictions.

[0074] S2. Based on the predicted sequence obtained in S1, construct an air conditioning load scheduling optimization model;

[0075] The objective function of the model includes two terms: net load cost and temperature / humidity over-limit penalty function, with the goal of minimizing the sum of the objective functions. The net load cost is calculated based on time-of-use electricity pricing and grid interaction power, where grid interaction power is the power consumption of the air conditioning system minus the photovoltaic power generation. The temperature / humidity over-limit penalty function penalizes the cumulative number of times the actual indoor temperature and humidity deviate from the set range. The optimization model uses the allowable indoor temperature and humidity setpoint range for the process environment as a hard constraint, and the predicted value of photovoltaic power generation as a constraint on real-time available power supply capacity. The allowable indoor temperature and humidity setpoint range for the process environment is a temperature of 25°C to 29°C and a relative humidity of 55%RH to 65%RH. In specific implementation, the constraints also include an upper limit on air conditioning power.

[0076] In practical implementation, the objective of the air conditioning load scheduling optimization model is:

[0077]

[0078]

[0079]

[0080] In the formula, The time-of-use electricity price for period t; and They represent the time to The cumulative number of times the indoor temperature or humidity exceeds the set range; N is the number of time periods; In time Net load In time Basic total load, In time air conditioning load, In time Photovoltaic power generation In time Electricity purchase price, In time The electricity price.

[0081] The penalty rules for failing to meet temperature and humidity standards are mathematically summarized as follows:

[0082]

[0083]

[0084] In the formula, This is the actual indoor temperature; and These are the upper and lower limits for setting the temperature; This refers to the actual indoor humidity. and It sets the upper and lower limits for humidity.

[0085] It should be noted that the mathematical formula of the temperature and humidity penalty function mentioned above is only an example. Those skilled in the art can use quadratic, exponential or other monotonically increasing function forms to achieve the same technical effect according to the technical requirements of "approaching the boundary penalty", and all of them fall within the protection scope of this invention.

[0086] In practical implementation, the decision variables of the optimization model include the dynamically changing power consumption of the air conditioning system, the indoor temperature setpoint, and the indoor relative humidity setpoint within a preset time period. Traditional methods typically only base scheduling on equipment start-up / shutdown or fixed temperature and humidity settings, while this scheme incorporates the temperature and humidity setpoints themselves as adjustable variables into the optimization framework, enabling the system to proactively "shape" the load curve within the allowable range of the process. This collaborative modeling approach significantly enhances the responsiveness of the scheduling strategy to fluctuations in photovoltaic power generation. Furthermore, by jointly optimizing power consumption and environmental setpoints, the algorithm can appropriately relax temperature and humidity settings (still within the process boundaries) during periods of low electricity prices or peak photovoltaic output to increase green electricity consumption or reduce electricity purchase costs; conversely, it can tighten settings during critical periods to ensure environmental quality. Compared to methods that only optimize power consumption or only adjust equipment status, this strategy more fully taps into the potential for flexible control.

[0087] In practical implementation, the relationship between the power consumption of the air conditioning system and the outdoor temperature and humidity and indoor temperature and humidity setpoints is determined by the following quantitative model of air conditioning load:

[0088] ;

[0089] In the formula, Let be the power consumption of the air conditioning system at time t; and These represent the outdoor temperature and relative humidity at time t, as predicted by S1. and These are the indoor temperature setting and indoor relative humidity setting at time t, respectively.

[0090] This approach enables precise modeling and improved predictability of air conditioning load. The model explicitly expresses air conditioning power consumption as a nonlinear function of outdoor temperature and humidity versus indoor setpoints, fully characterizing the impact of environmental parameter changes on energy consumption. Compared to empirical estimation or black-box simulation methods, this analytical model offers higher transparency and physical consistency, providing a reliable power consumption calculation basis for optimized scheduling. Furthermore, the model explicitly includes indoor temperature and humidity setpoints as variables, allowing for direct evaluation of the impact of different setting strategies on system power consumption during optimization. This supports the flexible control logic of "temperature-based load control," enhancing the scientific rigor and operability of scheduling decisions. This quantitative air conditioning load model provides a crucial mathematical tool for achieving source-load coordinated optimization, effectively improving the system's energy management accuracy and responsiveness under complex operating conditions.

[0091] S3. The improved market supervision optimization algorithm is used to solve the optimization model constructed in S2 to obtain the air conditioning load scheduling strategy and the corresponding dynamic temperature and humidity setpoints.

[0092] The improved market supervision optimization algorithm constructs a collaborative optimization framework encompassing four key roles: simulated business management, market transactions, fairness mechanisms, and feedback mechanisms. Enhancements are achieved within this framework through the following methods:

[0093] a) During the population exploration phase, the management role, market transaction role, fairness mechanism role and feedback mechanism role respectively execute differentiated search strategies, and adaptive random perturbation is introduced during the individual update process to enhance the global exploration capability;

[0094] b) During the local development phase, perform gradient-guided crossover and mutation operations on dominant individuals to improve convergence accuracy;

[0095] c) During the algorithm iteration process, the population diversity index is monitored in real time. When the index is lower than the preset threshold, the population dynamic recombination mechanism is triggered, and the feedback mechanism is used to guide the population out of the local optimum to prevent premature convergence.

[0096] In practice, in each iteration of S3, the search weight coefficients configured for the management role, market transaction role, fairness mechanism role, and feedback mechanism role are dynamically and adaptively adjusted based on the current iteration number and the rate of change of the population's historical best solution, in order to balance global exploration and local development capabilities.

[0097] This enhances the algorithm's adaptability at different optimization stages. The search weight coefficients are adjusted in real-time based on the current iteration progress and the rate of change of the population's historical best solutions. This allows the algorithm to focus more on global exploration in the early stages (e.g., dominating management and market transaction roles) and more on local fine-tuning in the later stages (e.g., strengthening fairness and feedback mechanisms). Compared to strategies with fixed weights or those relying solely on the number of iterations, this mechanism responds more sensitively to the actual state of the optimization process, improving search efficiency. Furthermore, by introducing the rate of change of historical best solutions as a feedback signal, the weight of exploratory roles is automatically increased when optimization stagnates, preventing premature convergence; while development capabilities are strengthened during periods of rapid performance improvement, accelerating convergence. This dynamic balancing mechanism is significantly superior to traditional static or heuristic weight allocation methods, improving the algorithm's overall robustness and solution quality.

[0098] In specific implementation, in step S3a), the random disturbance is achieved by superimposing noise that follows a Gaussian distribution onto the individual position vector, and the standard deviation of the Gaussian distribution adaptively decreases as the number of iterations increases.

[0099] Thus, in the early stages of the algorithm, the larger standard deviation of Gaussian noise enhances the intensity of random perturbations in the population, helping to escape local regions and expand the search range. As iterations progress, the standard deviation gradually decays, the perturbation amplitude decreases, and individuals more stably cluster towards the high-quality solution region. Compared to random perturbations of fixed intensity, this strategy more reasonably matches the needs of different stages of the optimization process. Furthermore, the adaptive decay mechanism avoids solution oscillations or convergence failures caused by excessively strong perturbations in the later stages, and also prevents the loss of diversity due to insufficient perturbations. Compared to traditional constant or linear perturbation methods, this method maintains population vitality while ensuring convergence accuracy, effectively mitigating premature convergence in complex multi-constraint scheduling problems.

[0100] In step S3b), the gradient guidance information is obtained by performing a finite difference approximation of the objective function in the neighborhood of the current dominant individual, and is used to guide the direction and step size of crossover and mutation operations.

[0101] Traditional evolutionary algorithms typically rely on random operations for crossover and mutation, lacking awareness of the local shape of the objective function. This scheme estimates the local gradient direction by performing finite difference approximation within the neighborhood of dominant individuals, guiding subsequent operations towards directions more likely to reduce operating costs, significantly enhancing the targeting and convergence speed of local optimization. Furthermore, industrial air conditioning scheduling models often contain non-differentiable or black-box components (such as equipment start-up / shutdown logic and discrete control variables), making it difficult to directly apply gradient-based optimization methods. This scheme implicitly obtains gradient information through finite difference, preserving the global framework of evolutionary algorithms while integrating the fine-tuning capabilities of gradient-based methods, balancing flexibility and accuracy.

[0102] In step S3c), the population diversity index is calculated using the average Euclidean distance between all individuals in the population, and the preset threshold is set to 10% to 50% of the population diversity index value when the algorithm is initialized.

[0103] In this way, the mean Euclidean distance effectively reflects the breadth of individual distribution in the solution space, and more accurately characterizes the spatial dispersion of the population compared to diversity indicators that rely solely on fitness variance or simple counting. Combining this with a relative threshold based on the initial diversity avoids the problem of poor adaptability of fixed thresholds across different problem sizes, making diversity assessment more robust and transferable. Furthermore, setting the threshold to a proportional range of the initial diversity (10%–50%) ensures that the algorithm initiates local development only after sufficient exploration, and also activates the feedback mechanism for dynamic reorganization when the population aggregates too early. Compared to absolute thresholds without threshold control or empirically set thresholds, this strategy better aligns with the actual evolutionary patterns of the optimization process, significantly improving the ability to escape local optima.

[0104] In specific implementation, in step S3c), when the population dynamic recombination mechanism is triggered, the best elite individuals are retained, and the remaining individuals are randomly initialized or perturbed based on the information of the elite individuals, so as to reconstruct the population and maintain global exploration capability.

[0105] In this way, retaining a few of the best elite individuals ensures that existing high-quality solutions are not discarded, avoiding the loss of optimization results; simultaneously, random or elite-information-based perturbation initialization of non-elite individuals effectively injects new solution space regions, quickly restoring population diversity. Compared to a full population reset or a completely random restart, this strategy achieves a better balance between exploration and exploitation. Furthermore, when premature convergence is detected in the population, this recombination mechanism can quickly break the clustering of individuals, guiding the search process to expand back into less explored regions. Especially in high-dimensional, multi-peak air conditioning load scheduling problems, this method significantly enhances the algorithm's ability to escape local traps, improving the stability and reliability of the final solution.

[0106] To facilitate understanding by those skilled in the art, the following explanation is provided.

[0107] The core concept of the original Market Regulation Optimization (MRO) algorithm is to simulate the four mechanisms of real-world market regulation: management (controlling the search step size through adjustment factor α), market transactions (introducing random perturbation β to enhance diversity), fairness mechanism (dynamic weight allocation using balance factor γ), and feedback mechanism (adjusting direction by correction factor δ). This constructs an optimization framework from initializing the population and global exploration to local development, aiming to solve complex optimization problems through multi-mechanism collaboration. The improved MRO (IMRO) further optimizes this framework: it strengthens multi-strategy parallel search in the exploration phase (such as guided search, fair search, and random perturbation combined), introduces fine-grained local search techniques such as gradient guidance and crossover mutation in the development phase, and adds a mechanism to prevent premature convergence (triggering dynamic recombination by real-time monitoring of population diversity). Ultimately, this significantly improves the algorithm's convergence accuracy, stability (e.g., the standard deviation of the test function approaches zero), and global search capability, making it more efficient for industrial optimization scenarios such as the scheduling of air conditioning systems in cigarette factories.

[0108] The process of using the algorithm (IMRO algorithm) in this invention is as follows: Figure 4 As shown, the IMRO algorithm incorporates three core improvements over the basic MRO algorithm, forming a more complete optimization framework:

[0109] Strategic level (business management): Dynamically balance exploration and development through an adaptive parameter adjustment mechanism.

[0110] Tactical layer (market trading): Introducing multi-strategy search patterns to enhance algorithm diversity.

[0111] Execution layer (fairness mechanism): Establish an elite retention and population recombination mechanism to prevent premature convergence.

[0112] Monitoring layer (feedback mechanism): Evaluates search performance in real time and dynamically adjusts search strategies.

[0113] S4. The air conditioning load scheduling strategy and dynamic temperature and humidity setpoints obtained in S3 are sent to the environmental control equipment for execution. By dynamically adjusting the temperature and humidity setpoints within the allowable range of the process, the air conditioning load curve is actively shaped so that it matches the fluctuation characteristics of photovoltaic power generation in the time dimension.

[0114] Figure 2 It clearly depicts the complete closed-loop process from data input (meteorological data, photovoltaic data) to final control execution. Figure 2 This highlights the core innovative idea of ​​the invention—predictive-based flexible closed-loop control. The process begins with prediction, then dynamically adjusts the setpoint, and uses an optimization algorithm to solve for the optimal scheduling command, ultimately achieving coordination between air conditioning load and photovoltaic power generation. The diagram clearly illustrates key steps such as "predicting photovoltaic power generation and outdoor temperature and humidity," "dynamically changing temperature and humidity setpoints," and "flexibly allocating air conditioning load power," intuitively demonstrating the essential difference from traditional open-loop or fixed-strategy control.

[0115] S5. Collect current actual meteorological data and photovoltaic power generation output data at preset time intervals, use them as inputs to update the latest historical data in S1, and re-execute steps S1 to S4 to form a dynamic rolling optimization closed loop, continuously ensuring the stability of the process environment and energy economy. The preset time interval is 15 minutes.

[0116] Compared with existing technologies, this invention introduces a dynamic temperature and humidity setting mechanism, allowing the system to proactively adjust the air conditioning load curve while meeting process constraints, thus better tracking the fluctuation characteristics of photovoltaic power generation over time. Compared to the rigid control strategy with fixed setpoints, this method significantly enhances the load-side response flexibility to photovoltaic output, effectively increasing the self-consumption ratio of photovoltaic power and reducing dependence on high-priced grid electricity. Furthermore, the Transformer-GRU hybrid prediction model integrates the global attention mechanism of Transformer with the local temporal modeling capability of GRU, and strengthens key feature extraction through an additive fusion mechanism. Compared to general prediction models with single structures (such as standard LSTM or pure Transformer), this model is more suitable for the complex data characteristics of strong coupling between photovoltaic output and micro-meteorological conditions in industrial scenarios such as cigarette factories, thus providing a more reliable input basis for subsequent optimization. Moreover, the improved market supervision optimization algorithm constructs a collaborative framework of four roles, introduces differentiated search strategies and adaptive perturbation mechanisms in the population exploration stage, combines gradient-guided information for refined operations in the local development stage, and is supplemented by population diversity monitoring and dynamic recombination mechanisms. Compared to traditional particle swarm optimization, genetic algorithms, or standard market surveillance algorithms, this method effectively alleviates the premature convergence problem and can still stably converge to a high-quality feasible solution under high-dimensional nonlinearity and multiple constraints. This method directly embeds the temperature and humidity setpoint range as a hard constraint into the optimization model, ensuring that all scheduling strategies strictly meet production process requirements. This avoids the problem of some existing economy-oriented scheduling methods becoming unexecutable due to neglecting hard constraints, significantly improving the engineering practicality and field deployability of the optimization results. Furthermore, by periodically updating actual operating data and re-executing the prediction-optimization-execution process, the system can dynamically adapt to uncertainties such as sudden weather changes, equipment aging, or load disturbances, forming a robust rolling scheduling mechanism. Compared to static or single-cycle optimization schemes, this closed-loop structure significantly enhances the long-term operational stability and energy management intelligence of the system.

[0117] This method is based on predictive flexible closed-loop control. The process begins with prediction, then dynamically adjusts the setpoint, and uses an optimization algorithm to solve for the optimal scheduling command, ultimately achieving coordination between air conditioning load and photovoltaic power generation. Experimental verification shows that this method has the following advantages compared to existing technologies:

[0118] 1. Improved forecast accuracy: The new Transformer-GRU model reduces the root mean square error (RMSE) and mean absolute percentage error (MAPE) of meteorological and photovoltaic forecasts by an average of 17.63%.

[0119] 2. Algorithm optimization: The improved market supervision optimization algorithm (IMRO) enhances diversity and global convergence, with the optimal value in the test function approaching 0 and the standard deviation being minimized.

[0120] 3. Reduced energy consumption: Air conditioning load decreased by 10.86%, and the proportion of photovoltaic use increased by 10.31%.

[0121] 4. Improved control accuracy: The overshoot rates for temperature and humidity decreased by 49.27% ​​and 47.97% respectively, with minimal fluctuations.

[0122] 5. Economic efficiency: By optimizing electricity purchase through time-of-use pricing, operating costs are significantly reduced.

[0123] Example 2

[0124] To help those skilled in the art better understand this method, the following examples are provided.

[0125] Example 1

[0126] Performance verification of the IMRO algorithm on the CEC test function

[0127] Example 1 verifies the superior performance of the IMRO algorithm (i.e., the improved market supervision optimization algorithm of this method) through standard test functions.

[0128] (1) Test environment

[0129] Hardware: Intel i7-12700H processor, 16GB RAM

[0130] Software: Python 3.8, NumPy scientific computing library

[0131] Test functions: Representative functions from the CEC test function set are selected, and several classic test functions are added to form a comprehensive test set covering characteristics such as unimodal, multimodal, and fixed-dimensional multimodal. Specifically, these include: Sphere, Quartic, Generalized Schwefel's Problem, Ackley, Hartman's Family, and Shekel's Foxholes functions.

[0132] Comparison Functions: To comprehensively evaluate performance, the Grey Wolf Optimization Algorithm (GWO), Whale Optimization Algorithm (WOA), and Basic Market Regulation Optimization Algorithm (MRO) were selected as comparison benchmarks.

[0133] (2) Parameter settings: Population size N=30, maximum number of iterations Tmax=1000, problem dimension D=30

[0134] (3) Test Result Analysis

[0135] Six test functions were selected and run independently 30 times. The results are shown in the table below:

[0136] Table 1 Test function results

[0137]

[0138] (4) Interpretation of results

[0139] a. Convergence accuracy analysis

[0140] In terms of convergence accuracy, the I MRO algorithm demonstrates superior performance on all test functions. On the unimodal function Sphere, the algorithm stably converges to the theoretical optimum, and the results of 30 independent runs are completely consistent, proving its extremely strong local search capability. On the multimodal function Ackley, the algorithm effectively escapes local optima and finds an optimal solution with near-machine accuracy, demonstrating excellent global exploration capability. Particularly on the complex fixed-dimensional multimodal function Shekel's Foxholes, the algorithm not only finds the theoretical optimum, but its average value is also infinitely close to the optimal solution, indicating that the algorithm has good stability when handling complex problems.

[0141] b. Stability and robustness analysis

[0142] In terms of stability, the IMRO algorithm performs exceptionally well. Looking at the standard deviation, the algorithm's standard deviation across all test functions is significantly lower than that of the comparison algorithms. Particularly on functions like Sphere and Ackley, the standard deviation reaches 0.0000e+00 or close to machine precision, indicating that the algorithm's results are almost unaffected by the randomness of the initial population, demonstrating excellent reproducibility. On the Quartic function, although all algorithms have relatively large standard deviations, IMRO's standard deviation is still significantly lower than MRO and other comparison algorithms, showing better robustness.

[0143] c. Comprehensive performance advantage analysis

[0144] Compared to mainstream algorithms such as GWO and WOA, IMRO demonstrates significant advantages across all metrics. In the six test functions, IMRO outperforms the benchmark in terms of optimal value, mean, and standard deviation. Compared to the basic MRO algorithm, IMRO shows significant improvements, particularly in the mean and standard deviation, validating the effectiveness of the algorithm improvements. This comprehensive performance advantage makes the IMRO algorithm particularly suitable for practical engineering problems requiring high stability and accuracy, such as the optimization of air conditioning systems in cigarette factories.

[0145] In summary, the experimental results demonstrate that the IMRO algorithm exhibits significant advantages in convergence accuracy, stability, and overall performance, providing a solid technical foundation for its practical application in industrial optimization problems. The algorithm's superior performance is primarily attributed to its unique market-monitoring mechanism, which effectively balances the exploration and development process, avoids premature convergence, and thus maintains stable high-performance output across various complex optimization problems.

[0146] Example 2

[0147] The specific implementation and effectiveness verification of the prediction model.

[0148] (1) Structural composition and working principle

[0149] Structure: The model input dimension is 24 (corresponding to 24 hours of data), containing 4 self-attention heads, each with 32 key channels. The model is based on the standard Transformer encoder-decoder structure.

[0150] (2) The key improvements are: embedding a GRU layer after the position coding layer to enhance temporal memory; adding an addition layer to merge encoder output and decoder information; introducing a dropout layer (Dropout=0.1) to prevent overfitting; and finally outputting the predicted value through a fully connected layer and a regression layer.

[0151] (3) Principle: During model training, the learning rate is 0.05, the regularization coefficient is 0.005, and the gradient clipping threshold is 10. It achieves high-precision prediction by analyzing the complex nonlinear relationship between historical meteorological data (temperature, humidity, irradiance) and photovoltaic output.

[0152] (4) Functions and effects

[0153] Function: It provides high-precision, forward-looking data input for the entire flexible scheduling system, serving as a "telescope" for optimization decisions.

[0154] Results: During the 5-day test period, the RMSE and MAPE values ​​of this model were lower than those of the traditional Transformer model for all predicted targets (outdoor temperature, humidity, and photovoltaic power generation). As shown in the comparison of prediction curves, its prediction results fit the actual value curves more closely, with an overall improvement in prediction performance of 17.63%, significantly reducing the uncertainty of subsequent optimization scheduling. (See Table 2 and...) Figure 5 As shown.

[0155] Table 2 Comparison of Predictive Performance

[0156]

[0157] Example 3

[0158] Application of IMRO algorithm in air conditioning system integration in cigarette factories.

[0159] Example 3 uses a tobacco curing warehouse in a cigarette factory as an application scenario to demonstrate the effectiveness of a complete temperature and humidity control system integrating the IMRO algorithm in actual operation.

[0160] (1) System structure composition

[0161] Hardware layer: The factory rooftop is equipped with a 1MW photovoltaic power generation system, a heating, ventilation and air conditioning (HVAC) system, temperature and humidity sensors throughout the workshops, a weather monitoring station, smart meters, and a central control server.

[0162] Software layer: Deploying the integrated control system of the present invention, comprising:

[0163] Prediction module: Employs the Transformer-GRU model to provide rolling forecasts of photovoltaic power generation and outdoor temperature and humidity for the next 24 hours.

[0164] IMRO Optimization Decision Module: Integrates the above-mentioned high-performance IMRO algorithm, serving as the "intelligent brain" for scheduling.

[0165] Control execution module: Receives optimization instructions and sends them to the air conditioning equipment.

[0166] (2) Working principle and process

[0167] The system operates in a closed loop of "prediction-optimization-execution-feedback":

[0168] Current forecast: At 0:00 each day, the forecast module generates an accurate forecast curve for the next 24 hours based on the latest data.

[0169] Optimized scheduling: The IMRO optimization module takes predicted data, time-of-use electricity prices (peak / flat / valley), and air conditioning load models as inputs, and uses "lowest operating cost" and "meeting temperature and humidity standards" as multiple objectives to quickly solve for the optimal air conditioning power consumption plan Pac(t) and dynamic temperature and humidity setpoints Tset(t) and Hset(t). The efficient convergence capability of the IMRO algorithm ensures that a high-quality solution can be completed within minutes.

[0170] Real-time control: The control module executes optimization commands to dynamically adjust the operating status of the air conditioning system.

[0171] Rolling correction: Fine-tuning is performed every 15 minutes based on the deviation between actual monitoring data and predictions to achieve adaptive optimization.

[0172] (3) Implementation effect

[0173] During the three-month (summer) trial run, the system achieved the following significant results:

[0174] Energy efficiency and economy: The total energy consumption of the air conditioning system was reduced by 10.86%; the average self-consumption rate of photovoltaic power generation reached 95.93%, which is more than 10 percentage points higher than the traditional fixed setpoint mode, greatly reducing the cost of purchasing electricity from the external grid.

[0175] Quality control: The accuracy of temperature and humidity control in the workshop has been greatly improved. The overshoot of temperature and humidity has decreased by 49.27% ​​and 47.97% respectively, and has been consistently maintained within the process requirements range (27±2℃, 60±5%), ensuring the quality of tobacco curing.

[0176] Algorithm performance: When solving a high-dimensional optimization problem with 124 decision variables (air conditioning power consumption and set value) every day, the IMRO algorithm demonstrates fast convergence and high stability as verified in Example 1, ensuring that the scheduling scheme can achieve near-global optimal performance every day, which is the key to achieving the above-mentioned energy-saving effect.

[0177] (5) Conclusion

[0178] Example 3 demonstrates that deep integration of high-performance IMRO optimization algorithms with predictive models and control execution modules can construct an efficient, stable, and automated industrial energy management system. This system successfully solves the synergistic challenge of efficient photovoltaic utilization and precise production environment protection, providing a replicable and highly effective energy-saving and carbon-reduction solution for process industries such as cigarette factories.

[0179] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit the technical solutions. Those skilled in the art should understand that any modifications or equivalent substitutions to the technical solutions of the present invention without departing from the spirit and scope of the present invention should be covered within the scope of the claims of the present invention.

Claims

1. An industrial air conditioning load scheduling method based on an improved market regulation optimization algorithm, characterized in that, The method comprises the following steps: S1, collecting historical meteorological data and historical photovoltaic power generation output data, and inputting the data into a pre-trained Transformer-GRU hybrid prediction model to obtain a prediction sequence of photovoltaic power generation power and outdoor environmental parameters in a future preset time length; wherein the Transformer-GRU hybrid prediction model embeds a gated recurrent unit (GRU) layer in a standard Transformer encoder-decoder architecture, and enhances the time series feature extraction capability through an additive fusion mechanism; S2, based on the prediction sequence obtained in S1, an air conditioner load scheduling optimization model is constructed; the objective function of the model includes two items of net load cost and temperature and humidity overrun penalty function, and the sum of the objective functions is minimized as the goal; wherein the net load cost is calculated according to the time-of-use electricity price and the grid interactive power, and the grid interactive power is the air conditioning system power consumption minus the photovoltaic power generation power; the temperature and humidity overrun penalty function is used to punish the cumulative number of indoor actual temperature and humidity deviating from the set range; the optimization model takes the indoor temperature and humidity set value range allowed by the process environment as a hard constraint, and takes the prediction value of the photovoltaic power generation power as a real-time available power supply capacity constraint; S3, an improved market supervision optimization algorithm is used to solve the optimization model constructed in S2 to obtain an air conditioner load scheduling strategy and corresponding dynamic temperature and humidity set values; Wherein, the improved market supervision optimization algorithm constructs a collaborative optimization framework simulating four roles of management, market transaction, fairness mechanism and feedback mechanism, and enhances it in the following ways under the framework: a) In the population exploration stage, the management role, market transaction role, fairness mechanism role and feedback mechanism role respectively execute differentiated search strategies, and introduce adaptive random disturbance in the individual update process to enhance the global exploration ability; b) In the local development stage, the superior individuals are subjected to cross and mutation operations guided by gradient-oriented information to improve the convergence accuracy; c) In the algorithm iteration process, the population diversity index is monitored in real time, and when the index is lower than a preset threshold, a population dynamic restructuring mechanism is triggered, and the population is guided to jump out of the local optimum through the feedback mechanism to prevent premature convergence; S4, the air conditioner load scheduling strategy and dynamic temperature and humidity set values obtained in S3 are sent to the environmental control equipment for execution, and by dynamically adjusting the temperature and humidity set values within the process allowed range, the air conditioner load curve is actively shaped to match the fluctuation characteristics of the photovoltaic power generation power in the time dimension; S5, the current actual meteorological data and photovoltaic power generation output data are collected at a preset time interval as the latest historical data and updated to the input in S1, and the steps of S1 to S4 are re-executed to form a dynamic rolling optimization closed loop, continuously ensuring the process environment stability and energy economy.

2. The method for industrial air conditioning load scheduling based on improved market supervision optimization algorithm according to claim 1, characterized in that: In S2, the decision variables of the optimization model include the dynamically changing air conditioning system power consumption, indoor temperature set value and indoor relative humidity set value in each period within the future preset time length.

3. The method for industrial air conditioning load scheduling based on improved market supervision optimization algorithm according to claim 1, characterized in that: In S2, the relationship between the air conditioning system power consumption and the outdoor temperature and humidity and the indoor temperature and humidity set values is determined by the following air conditioning load quantitative model: ； In the formula, P(t) is the power consumption of the air conditioning system at time t; and Tout(t) and RHout(t) are the outdoor temperature and relative humidity at time t predicted by S1, respectively; and Tset(t) and RHset(t) are the indoor temperature setpoint and indoor relative humidity setpoint at time t, respectively.

4. The method for industrial air conditioning load scheduling based on improved market supervision optimization algorithm of claim 1, wherein: In each iteration of S3, the search weight coefficients configured for the management role, the market transaction role, the fairness mechanism role and the feedback mechanism role are dynamically self-adaptively adjusted according to the current iteration number and the change rate of the historical optimal solution of the population, so as to balance the global exploration and local development capabilities.

5. The method for industrial air conditioning load scheduling based on improved market regulation optimization algorithm of claim 1, wherein: In step S3a), the random disturbance is realized by superimposing noise obeying a Gaussian distribution on the individual position vector, and the standard deviation of the Gaussian distribution is self-adaptively attenuated with the increase of the iteration number.

6. The method for industrial air conditioning load scheduling based on improved market regulation optimization algorithm of claim 1, wherein: In step S3b), the gradient guidance information is obtained by performing a finite difference approximation of the objective function in the neighborhood of the current dominant individual, and is used to guide the direction and step length of the crossover and mutation operations.

7. The method for industrial air conditioning load scheduling based on improved market regulation optimization algorithm of claim 1, wherein: In step S3c), the population diversity index is calculated by using the average Euclidean distance between all pairs of individuals in the population, and the preset threshold is set to 10% to 50% of the population diversity index value at the initialization of the algorithm.

8. The industrial air conditioning load scheduling method based on improved market regulation optimization algorithm according to claim 1 or 7, characterized in that: In step S3c), when the population dynamic recombination mechanism is triggered, a number of elite individuals are retained, and the remaining individuals are randomly initialized or disturbed based on the information of the elite individuals to reconstruct the population and maintain the global exploration capability.

9. The method for industrial air conditioning load scheduling based on improved market regulation optimization algorithm of claim 1, wherein: In S1, in the Transformer-GRU hybrid prediction model, the GRU layer is embedded after the input position encoding of the Transformer encoder; the additive fusion mechanism is realized through an additive fusion layer, which is used to fuse the output features of the Transformer encoder and the intermediate feature representation of the decoder; the Transformer-GRU hybrid prediction model further includes a Dropout regularization layer and a fully connected regression output layer, wherein the dropout rate of the Dropout regularization layer is set to 0.

1.

10. The method for industrial air conditioning load scheduling based on improved market regulation optimization algorithm as claimed in claim 1, wherein: In S2, the indoor temperature and humidity setting value range allowed by the process environment is 25-29℃ for temperature and 55-65% RH for relative humidity.

Images