AI evaluation method for online load regulation capability of freezing plant

Abstract

The invention discloses a freezing plant load on-line regulation capability AI evaluation method, and belongs to the field of power systems. According to the method, an up-regulation power variation prediction model based on LSTM is constructed for a refrigeration compressor unit, dynamic evolution characteristics are extracted by using continuous time sequence characteristics, and the maximum rapid up-regulation capability under the current working condition is output; an up-regulation characteristic index prediction model based on XGBoost is constructed for the condenser fan, and the real-time regulation capability is evaluated through logarithmic transformation and nonlinear regression modeling of the relationship between the load rate and the power; a LightGBM-based rapid up-regulation coefficient prediction model is constructed for the evaporator fan of the refrigeration house, mapping between characteristics and adjustment response is carried out by combining the environmental characteristics of the refrigeration house, and the adjustment capacity of the evaporator fan is estimated; and finally, carrying out overall online adjustment capability AI evaluation, outputting the comprehensive adjustment capability at each time point, and carrying out refined modeling and real-time prediction on the adjustment potentials in the operation process of various key equipment of the freezing plant.

Description

Technical Field

[0001] The present invention relates to the field of power systems, and in particular to an AI evaluation method for online load regulation capability of a refrigeration plant. Background Art

[0002] Evaluating the online load regulation capabilities of refrigeration plants helps provide real-time insights into their adjustable power range under different operating conditions, enhancing their ability to participate as a flexible resource on the load side in grid peak regulation and demand response. Dynamic evaluation can optimize load control strategies, reduce operating costs, enhance system stability, and provide strong support for building a power system with a high proportion of renewable energy.

[0003] However, the existing methods for evaluating the online regulation capacity of refrigeration plants have the following major deficiencies: 1) Lack of equipment-level modeling: Most methods treat the refrigeration plant as an overall load unit, without breaking it down into key equipment such as compressors, condenser fans, and evaporator fans, making it difficult to accurately evaluate the regulation potential of each subsystem. 2) Ignoring differences in operating characteristics: Different equipment has different response characteristics to external factors such as temperature, humidity, and operating frequency, and most existing methods do not distinguish between them, resulting in deviations in the estimation of regulation capacity. 3) Insufficient modeling capabilities: Traditional empirical models and linear methods cannot capture the complex nonlinear relationship between the operating status of the equipment and external disturbances, and the prediction accuracy is limited. 4) Weak timeliness and generalization capabilities: Most methods find it difficult to achieve adaptive identification and generalization of regulation capabilities at different time scales and under multiple operating conditions, affecting the effectiveness of scheduling applications.

[0004] Therefore, it is necessary to combine advanced AI algorithms to establish an equipment-level evaluation model, conduct AI evaluation of online adjustment capabilities, accurately characterize the adjustable capabilities of key electrical equipment in refrigeration plants under different working conditions, and improve the dynamics and accuracy of adjustment capability evaluation. Summary of the Invention

[0005] The problem to be solved by the present invention is: to provide an AI evaluation method for the online load regulation capability of a refrigeration plant, combine the LSTM, XGBoost and LightGBM algorithms, establish an equipment-level evaluation model for refrigeration compressors, condenser fans, and evaporator fans, and integrate and construct an overall AI evaluation framework for the refrigeration plant, so as to improve the dynamics and accuracy of the regulation capability evaluation and provide intelligent decision-making support for its participation in grid peak regulation and demand response.

[0006] The present invention adopts the following technical solution: an AI evaluation method for online load regulation capability of a refrigeration plant, comprising the following steps:

[0007] Step 1: Build an LSTM-based model for evaluating the online adjustment capability of refrigeration compressor units in refrigeration plants. This model uses an LSTM-based model to predict the maximum allowable increase in compressor power. This model uses compressor operating data from continuous time windows and inputs it into an LSTM network to extract temporal dynamic features. This model then performs feature compression mapping and predicts the maximum allowable increase in compressor power at the current moment, thereby evaluating the online adjustment capability of the refrigeration compressor.

[0008] Step 2: Build an XGBoost-based model to evaluate the online regulation capability of condenser fans in refrigeration plants. This model uses an XGBoost-based condenser fan regulation index prediction model to predict the regulation index of the condenser fan load rate and power relationship, and evaluate the online regulation capability of the condenser fan. This model also evaluates the online regulation capability of the condenser fan by capturing the nonlinear response of fan load changes to power changes.

[0009] Step 3: Construct a LightGBM-based online adjustment capability evaluation model for the evaporator fan in the cold storage of the freezing plant: The LightGBM-based evaporator fan rapid increase coefficient prediction model is used to predict the rapid increase coefficient of the evaporator fan under specific operating conditions. The online adjustment capability of the cold storage evaporator fan is evaluated by comprehensively considering the impact of changes in the cold storage temperature control environment.

[0010] Step 4. Build an AI evaluation model for the online regulation capability of the refrigeration plant: Based on the evaluated online regulation capability of the refrigeration compressor unit, the online regulation capability of the condenser fan, and the online regulation capability of the cold storage evaporator fan, the online regulation capability of the refrigeration plant is evaluated to provide support for online grid regulation.

[0011] Preferably, in step 1, a LSTM-based prediction model for the maximum allowable increase in compressor power change is constructed, including the following sub-steps:

[0012] Step 101, feature collection within the time window: collecting compressor operating features for k consecutive time steps, including compressor power, head temperature, refrigerant inlet pressure and outlet pressure, to form an original feature matrix;

[0013] Step 102: Feature normalization: normalize the feature matrix X j (d,τ) is normalized to obtain the normalized compressor characteristic matrix

[0014] Step 103, LSTM dynamic feature encoding: Input the compressor feature matrix obtained by normalization into the LSTM network, extract the dynamic evolution characteristics of the system operation through time step recursion, and obtain the hidden state at each moment;

[0015] Step 104: Feature compression mapping: compress the hidden state output by the LSTM into an intermediate prediction vector through a fully connected layer.

[0016] Step 105 , predicting the maximum allowable power increase change: the intermediate feature vector is mapped to a scalar through another fully connected mapping and output, so as to predict the maximum power increase capability of the compressor.

[0017] Preferably, in step 1, the online adjustment capability of the compressor is evaluated, and the compressor power and load are adjusted based on the dynamic mode of load, temperature and pressure changes of the compressor in a short period of time, and the rapid increase capability is limited by the pressure fluctuation protection of the refrigeration system, which is expressed as follows:

[0018]

[0019] Where: C up,comp (d,t),C down,comp (d, t) is the compressor's adjustment capacity and reduction capacity at time t on day d, η comp,max is the maximum allowable load rate of the compressor, η comp (d,t) is the efficiency of the compressor at time t on day d, P comp,rated is the rated power of the compressor, ΔP comp,max is the maximum allowable increase in compressor power, η comp,min It is the minimum allowable load rate of the compressor.

[0020] Preferably, in step 2, a condenser fan upward regulation characteristic index prediction model based on XGBoost is constructed, and the specific steps are as follows:

[0021] Step 201, raw data collection: collecting the load rate and power data points of the condenser fan over a period of time as the raw data set D;

[0022] Step 202, feature transformation: perform logarithmic transformation on the original data set D, linearize the power function relationship, and obtain the logarithmic transformed data set D;

[0023] Step 203: Organize the training features and label sets: organize the logarithmically transformed dataset D into feature vectors X and label vectors Y respectively;

[0024] Step 204: Train the XGBoost regression model based on the feature set: train the XGBoost regression model to learn the mapping relationship between load rate and power, and use the trained regression function f XGB (·) output;

[0025] Step 205, local derivative calculation: During prediction, the trained regression function is differentiated near the current load rate, the power exponent is extracted, and the upward characteristic index γ of the wind turbine is obtained.

[0026] Preferably, in step 2, the online adjustment capability of the condenser fan is evaluated. The power of the condenser fan is approximately in a square relationship with the wind speed. The upward adjustment capability increases nonlinearly with the remaining wind speed space, and the downward adjustment capability increases linearly with the remaining wind speed space, which can be expressed as:

[0027]

[0028] Where: C up,cond (d,t),C down,cond (d, t) is the upward and downward adjustment capabilities of the condenser fan, η cond (d,t) is the original load rate of the condenser fan, P cond,rated is the rated power of the condenser fan, and γ is the fan sharing function adjustment index.

[0029] Preferably, in step 3, in the evaporator fan rapid increase coefficient prediction model based on LightGBM, the influencing factors of the evaporator fan rapid increase coefficient include: initial load rate, temperature change rate in the cold storage, humidity change rate in the cold storage, and cold storage environment capacity. By collecting the evaporator fan operation characteristics, the evaporator fan rapid increase coefficient is predicted, and the specific processing is as follows:

[0030] Step 301, data collection: collect the evaporator fan operation data at time t on day d, including: the current load rate of the evaporator fan, the temperature change rate after the evaporator fan is adjusted up, and the humidity change rate after the evaporator fan is adjusted up, and save the data to the feature vector x evap (d,t);

[0031] Step 302: Feature normalization: normalize each feature to obtain a scalarized feature vector.

[0032] Step 303: Organize the training data set: Organize all samples collected over a period of time into a training set and output the training feature set X evap and training label set Y evap ;

[0033] Step 304, LightGBM regression training: Based on the training set, train the LightGBM regressor, learn the mapping relationship from features to up-adjusted coefficients, and obtain the trained LightGBM regressor f LightGBM (·);

[0034] Step 305: Predict the evaporator fan rapid increase coefficient: Under the new operating state, input the normalized characteristics and predict the evaporator fan rapid increase coefficient θ evap (d,t).

[0035] Preferably, in step 3, the online adjustment capacity evaluation model of the evaporator fan in the cold storage of the freezing plant based on LightGBM is used to adjust the load through the start-stop group control, and the upward and downward adjustment capacity of the evaporator fan is evaluated:

[0036]

[0037] Where: C up,evap (d,t),C down,evap (d, t) is the evaporator fan's upward and downward adjustment capabilities, θ evap is the evaporator fan rapid adjustment coefficient, ζ evap is the evaporator fan rapid down-regulation coefficient, P evap,rated is the rated power of the evaporator fan.

[0038] Preferably, in step 4, an AI evaluation model for the online regulation capability of the refrigeration plant is constructed to evaluate the online regulation capability of the refrigeration plant. The formula is as follows:

[0039]

[0040] Where: C up,total (t), C down,total (t) is the overall power increase and power decrease, C up,comp (t), C down,comp (t) is the upward and downward adjustment capabilities of the compressor, C up,cond (t), C down,cond (t) is the upward and downward adjustment capabilities of the condenser, C up,evap (t), C down,evap (t) is the upward and downward adjustment capabilities of the evaporator.

[0041] The technical solution of the present invention further provides: an electronic device, comprising:

[0042] one or more processors;

[0043] a storage device having one or more programs stored thereon;

[0044] When the one or more programs are executed by the one or more processors, the one or more processors implement any of the above-mentioned AI evaluation methods for online load regulation capability of a refrigeration plant.

[0045] The technical solution of the present invention also provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the steps in any of the above-mentioned AI evaluation methods for online load regulation capability of a refrigeration plant are implemented.

[0046] Compared with the prior art, the present invention adopts the above technical solution and has the following technical effects:

[0047] 1. The evaluation method of the present invention uses LSTM to model the time series of the operating power of the compressor unit, extracts its dynamic response characteristics, and realizes accurate prediction of future regulation capacity; combines multi-dimensional inputs such as ambient temperature and equipment load, and uses XGBoost to model its nonlinear regulation capacity boundary to improve prediction accuracy; in view of the frequent start and stop and power fluctuation characteristics of the evaporator fan, a fast training and high-performance prediction model is constructed based on the LightGBM algorithm to predict the regulation capacity of the evaporator fan.

[0048] 3. The evaluation method of the present invention establishes an equipment-level evaluation model for refrigeration compressors, condenser fans, and evaporator fans, accurately depicting the adjustable capabilities of key electrical equipment in refrigeration plants under different operating conditions, improving the dynamics and accuracy of adjustment capability evaluation, and integrating the construction of an overall AI evaluation framework for refrigeration plants, providing intelligent decision-making support for them to participate as adjustable loads in grid peak regulation and demand response. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 This is a flowchart of the AI ​​evaluation method for online load regulation capability of a refrigeration plant according to the present invention. DETAILED DESCRIPTION

[0050] In order to make the purpose, technical solutions and advantages of the present invention clearer, the technical solutions of the application are further elaborated in detail below with reference to the accompanying drawings. The described embodiments are only a part of the embodiments involved in the present invention. All non-innovative embodiments of other researchers in this field on this embodiment fall within the scope of protection of the present invention. At the same time, the step numbers in the embodiments of the present invention are only set for the convenience of explanation and description, and the order between the steps is not limited in any way. The execution order of each step in the embodiment can be adaptively adjusted according to the understanding of those skilled in the art.

[0051] In one embodiment of the present invention, an AI evaluation method for the online load regulation capability of a freezing plant mainly includes: an online regulation capability evaluation model of the refrigeration compressor unit of the freezing plant based on LSTM, an online regulation capability evaluation model of the condenser fan of the freezing plant based on XGBoost, an online regulation capability evaluation model of the evaporator fan of the cold storage of the freezing plant based on LightGBM, and an AI evaluation model for the online regulation capability of the freezing plant.

[0052] like Figure 1 As shown, the AI ​​evaluation method for online load regulation capability of a refrigeration plant in this embodiment has the following specific steps:

[0053] Step 1: Build an LSTM-based online regulation capability evaluation model for refrigeration compressor units in refrigeration plants.

[0054] In this embodiment, the LSTM-based prediction model for the maximum allowable increase in compressor power change is used to predict the maximum power change that the compressor can quickly and safely increase under the current operating state. The solution process is as follows:

[0055] First, the compressor operation characteristics of k consecutive time steps are collected, including compressor power, head temperature, refrigerant inlet pressure and outlet pressure, to form the original feature sequence.

[0056] Each feature is then normalized to eliminate the effects of dimensional differences. The normalized feature sequence is fed into the LSTM network as input, where it extracts the dynamic evolution of the system's operation through time-step recursion. The hidden state vector at the final moment of the LSTM serves as a compressed representation of the overall time series features and is further mapped into an intermediate feature vector through a fully connected layer.

[0057] Then, the intermediate feature vector is mapped through another fully connected map to output the maximum allowable rapid increase in power change of the compressor at the current moment.

[0058] Finally, an online adjustment capability evaluation model for the refrigeration compressor unit is developed. The maximum allowable increase in power change of the refrigeration compressor unit given by LSTM is used to construct a model to evaluate the online adjustment capability of the refrigeration compressor unit in the freezing plant.

[0059] The method of this embodiment can capture the dynamic pattern of load, temperature, and pressure changes of the compressor in a short period of time, and provide continuous and real-time prediction of safe increase capacity.

[0060] Step 2: Propose an online regulation capability evaluation model for condenser fans in refrigeration plants based on XGBoost.

[0061] In this embodiment, the condenser fan upward characteristic index prediction model based on XGBoost is used to predict the upward characteristic index of the relationship between the condenser fan load rate and power. The solution process is as follows:

[0062] First, the original data points of the load rate and corresponding power of the condenser fan under different operating conditions are collected. In order to linearize the power function relationship between the fan power and the load rate, the logarithms of the load rate and power are taken respectively to obtain the linearized characteristic input and output.

[0063] Next, all logarithmized samples are organized into feature sets and label sets. An XGBoost regression model is then trained to learn the nonlinear mapping between features and labels. After training, the wind turbine's ramp-up characteristic index for the current operating point is obtained by taking the local derivative of the regression model in the input space and extracting the slope in logarithmic space.

[0064] Finally, the online regulation capability of the condenser fan is evaluated. Based on the condenser fan upward regulation characteristic index given by XGBoost, a model is constructed to evaluate the online regulation capability of the condenser fan.

[0065] The method of this embodiment can accurately capture the nonlinear response law of wind turbine load changes to power changes, and supports flexible and accurate rapid adjustment capability modeling.

[0066] Step 3: Propose an online adjustment capability evaluation model for the evaporator fan of the cold storage in the freezing plant based on LightGBM.

[0067] In this embodiment, the evaporator fan rapid increase coefficient prediction model based on LightGBM is used to predict the rapid increase coefficient of the evaporator fan under specific operating conditions. The solution process is as follows:

[0068] First, the real-time operating characteristics of the evaporator fan are collected, including the initial load rate, the rate of change of the cold storage temperature and the rate of change of the humidity after the increase, to form a single-step feature vector. To ensure numerical stability and training efficiency, each feature is standardized.

[0069] Subsequently, the collected series of standardized features and the corresponding historical observation up-adjustment coefficients are combined into a training sample set, which is trained using the LightGBM regressor to establish a nonlinear mapping relationship between features and fast up-adjustment coefficients.

[0070] In practical applications, by inputting the current standardized features, the corresponding evaporator fan rapid adjustment coefficient can be predicted through the trained LightGBM model.

[0071] Finally, the online adjustment capability of the cold storage evaporator fan is evaluated. Based on the rapid increase coefficient of the evaporator fan predicted by LightGBM, a model is constructed to evaluate the online adjustment capability of the cold storage evaporator fan.

[0072] The method of this embodiment can comprehensively consider the impact of changes in the cold storage temperature control environment and provide an accurate real-time estimation of the evaporator fan's rapid load response capability.

[0073] Step 4: Propose an AI evaluation model for the online adjustment capability of the refrigeration plant.

[0074] In this embodiment, the online adjustment capability evaluation model of the refrigeration compressor unit of the freezing plant based on LSTM, the online adjustment capability evaluation model of the refrigeration plant condenser fan based on XGBoost, and the online adjustment capability evaluation model of the cold storage evaporator fan of the freezing plant based on LightGBM are used to obtain the online adjustment capability of the refrigeration compressor unit, the online adjustment capability of the condenser fan, and the online adjustment capability of the cold storage evaporator fan. A model is constructed to evaluate the online adjustment capability of the freezing plant to provide support for the online regulation of the power grid.

[0075] Specifically, for step 1, this embodiment proposes an LSTM-based prediction model for the maximum allowable increase in compressor power change. Using compressor operating data (power, temperature, pressure, etc.) in a continuous time window, the LSTM is used to extract temporal dynamic features. After feature compression mapping, the maximum allowable increase in power change at the current moment is predicted. The specific process is as follows:

[0076] (1) Feature collection within the time window

[0077] Define that at time t on day d, the compressor characteristics of the most recent k steps are collected, including compressor power, compressor temperature, and compressor pressure. The formula is as follows:

[0078] X comp (d,t)={[P comp (d,τ),T comp (d,τ),P refrig.in (d,τ),P refrig.out (d,τ)]|τ=t-k+1,...,t}

[0079] Where: X comp (d, t) is the compressor characteristic sequence matrix, P comp (d,τ) is the power of the compressor at the time τ on the dth day, T comp (d,τ) is the temperature of the compressor at the time τ on the dth day, P refrig.in (d,τ) is the compressor input pressure at the dth day and time period τ, P refrig.out (d,τ) is the compressor output pressure at time τ on day d.

[0080] (2) Feature normalization processing

[0081] In order to eliminate the influence of features of different dimensions on training, the compressor feature sequence matrix is ​​normalized by column (standardized to mean 0 and variance 1). The normalization operation is as follows:

[0082]

[0083] The normalized result is as follows:

[0084] X comp (d,t)=Norm(X comp (d,t))

[0085] Where: is the normalized feature matrix, X j (d,τ) is the feature matrix before normalization, μ j is the mean of the j-th class feature, σ j is the standard deviation of the j-th class feature, Xcomp (d, t) is the characteristic matrix after the compressor is normalized, X comp (d,t) is the characteristic sequence matrix of the compressor before normalization.

[0086] (3) LSTM dynamic feature encoding

[0087] The normalized compressor feature matrix is ​​input into the LSTM network to obtain the hidden state at each moment. The formula is as follows:

[0088]

[0089] Where: h τ is the hidden state vector at the τth moment, c τ is the memory state of the unit at the τth moment, LSTMCell(·) is the LSTM network, X comp (d,τ) is the normalized compressor characteristic matrix, h τ-1 is the hidden state vector at the τ-1th moment, c τ-1 is the unit memory state at the τ-1th moment, h comp (d, t) is the hidden state vector of LSTM at the last moment.

[0090] (4) Feature compression mapping

[0091] The hidden layer state of the LSTM output is compressed and mapped into an intermediate prediction vector through a fully connected layer. The formula is as follows:

[0092] z comp (d,t)=W z ·h comp (d,t)+b z

[0093] Where: z comp (d, t) is the intermediate feature compression vector, W z is the feature map weight matrix, b z is the feature map bias vector, h comp (d, t) is the hidden state vector of LSTM at the last moment.

[0094] (5) Prediction of maximum allowable power increase change

[0095] The intermediate feature vector is mapped to a scalar and output to predict the maximum increase in the compressor capacity, where the predicted increase in the compressor capacity is expressed as ΔP comp,max (d,t):

[0096]

[0097] Where: ΔP comp,max(d, t) is the maximum allowable rapid increase in power change (kW) predicted in the final forecast, w out is the output weight vector, b out is the output bias term, z comp (d,t) is the intermediate feature compression vector.

[0098] Furthermore, in the online adjustment capability evaluation model of the refrigeration compressor unit based on LSTM in this embodiment, power regulation is usually achieved through variable frequency control (continuously adjustable within a certain range), the load adjustment range is usually the partial load area (50% to 100%), and the rapid adjustment capability is limited by the pressure fluctuation protection of the refrigeration system.

[0099] Since the compressor's up / down adjustment capability can only exist within the operational load range, its up / down adjustment capability will be limited and cannot be increased arbitrarily. The compressor's up / down adjustment capability is calculated as follows:

[0100]

[0101] Where: C up,comp (d,t) is the compressor's adjustment capacity at time t on day d, η comp,max is the maximum allowable load rate of the compressor, η comp (d,t) is the efficiency of the compressor at time t on day d, P comp,rated is the rated power of the compressor, ΔP comp,max The maximum allowable increase in compressor power, C down,comp (d,t) is the compressor's down-regulation capacity at time t on day d, η comp,min The minimum allowable load rate of the compressor.

[0102] Specifically, for step 2, this embodiment proposes a condenser fan upward characteristic index prediction model based on XGBoost, using the actual power data of the condenser fan under different load rates, through feature transformation (logarithm), the power function relationship is linearized, and the XGBoost regression model is trained to fit the relationship between load rate and power. The fan's upward characteristic index γ is inferred through the local derivative of the regression model.

[0103] (1) Collecting original data of fan operation

[0104] Collect the load rate and power data points of the fan over a period of time and use the data as the original data set D:

[0105]

[0106] Where: D is the original data set, recording the load rate and power, is the sampling time set, η cond(d, t) is the working efficiency of the condenser at time t on day d, P cond (d,t) is the power consumption of the condenser at time t on day d.

[0107] (2) Logarithmic transformation of original data

[0108] To linearize the power function relationship P∝η γ , perform logarithmic transformation on the data to obtain the logarithmic transformed data set D:

[0109]

[0110] Where: is the working efficiency of the condenser at time t on day d after logarithmic transformation, is the power consumption of the condenser at time t on day d after logarithmic transformation, η cond (d, t) is the working efficiency of the condenser at time t on day d before logarithmic transformation, P cond (d, t) is the power consumption of the condenser at time t on day d before logarithmic transformation; D is the new data set obtained by logarithmic transformation of the original data set D; log(·) is the logarithmic function.

[0111] (3) Organizing training features and label sets

[0112] Organize the logarithmic data into feature vector X and label vector Y respectively:

[0113]

[0114] Where: is the working efficiency of the condenser at time t on day d after logarithmic transformation, is the electric power corresponding to the condenser at time t on day d after logarithmic transformation.

[0115] (4) Training XGBoost regression model based on feature set

[0116] Train the XGBoost regression model, learn the mapping relationship between load rate and power (logarithmic space), and use the trained regression function f XGB (·) to output:

[0117]

[0118] Where: f XGB (·) is the trained regression function, is the electric power corresponding to the condenser at time t on day d after logarithmic transformation, is the working efficiency of the condenser at time t on day d after logarithmic transformation, f XGB(·) is the trained XGBoost regressor and ε(d,t) is the fitted residual noise term.

[0119] (5) Local derivative calculation: extracting the power exponent γ

[0120] During prediction, the trained regressor is differentiated around the current load rate to obtain the up-regulation characteristic index:

[0121]

[0122] Where: γ(d,t) is the condenser fan upward characteristic index at time t on day d, is the local first-order derivative of the regression function.

[0123] Furthermore, in the XGBoost-based condenser fan online regulation capability assessment model proposed in this embodiment, fans generally use multi-stage variable speed or variable frequency speed regulation, which minimizes the impact of load changes on the system and allows for rapid start / stop or speed changes. Increased speed of the condenser fan is primarily limited to its maximum speed, while decreased speed can be close to a stop.

[0124] Since the condenser fan power is approximately in a square relationship with the wind speed, its upward adjustment capacity increases nonlinearly with the remaining wind speed space. However, the downward adjustment of the condenser fan is relatively simple and linear. The upward adjustment is affected by nonlinearity and must be corrected. The upward and downward adjustment capacity of the condenser fan are calculated as follows:

[0125]

[0126] Where: C up,cond (d, t) is the upward adjustment capacity of the condenser fan, C down,cond (d,t) is the down regulation capability of the condenser fan, η cond (d,t) is the original load rate of the condenser fan, P cond,rated is the rated power of the condenser fan and the γ fan sharing function adjustment index (through empirical fitting).

[0127] Specifically, for step 3, the evaporator fan rapid increase coefficient prediction model based on LightGBM proposed in this embodiment, the evaporator fan rapid increase coefficient is affected by the following factors: initial load rate, temperature change rate in the cold storage, humidity change rate in the cold storage, cold storage environment capacity, etc. By collecting operating characteristics, the LightGBM regression model is used to learn the mapping relationship between characteristics and safety increase coefficients.

[0128] (1) Collecting evaporator fan operation data

[0129] At time t on day d, the following features are collected: the current load rate of the evaporator fan, the temperature change rate after the evaporator fan is adjusted upward, and the humidity change rate after the evaporator fan is adjusted upward, and the features are saved in the feature vector:

[0130] x evap (d,t)=[η evap (d,t),ΔT evap (d,t),ΔH evap (d,t)]

[0131] Where: x evap (d,t) is the eigenvector at a single moment, η evap (d, t) is the current load rate of the evaporator fan, ΔT evap (d, t) is the temperature change rate after the evaporator fan is adjusted upward (℃ / min), ΔH evap (d, t) is the humidity change rate after the evaporator fan is adjusted upward (% / min).

[0132] (2) Feature normalization processing

[0133] Standardize each type of feature to prevent dimensional differences from affecting model training:

[0134]

[0135] Where: is the scalarized eigenvector, x j (d,t) is the eigenvector before scalarization, μ j is the mean of the j-th feature, σ j is the standard deviation of the j-th class feature.

[0136] (3) Organizing training datasets

[0137] Organize all samples collected over a period of time into a training set and output the training feature set X evap and training label set Y evap :

[0138]

[0139] Where: X evap is the training feature set, is the normalized eigenvector, Y evap is the training label set, θ evap,obs (d,t) is the rapid upward adjustment coefficient obtained from historical experiments or empirical observations.

[0140] (4) LightGBM regression training

[0141] Based on the training set, train the LightGBM regressor to learn the mapping relationship between features and up-regulation coefficients, which is expressed as follows:

[0142] θ evap,obs (d,t)=f LightGBM (x evap (d,t))+ε(d,t)

[0143] Where: θ evap,ob ( s d,t) is the rapid upward adjustment coefficient obtained from historical experiments or empirical observations, f LightGBM (·) is the trained LightGBM regressor, ε(d,t) is the residual error, is the normalized feature vector.

[0144] (5) Predict the evaporator fan rapid adjustment coefficient

[0145] In the new operating state, input the normalized features and predict the rapid increase coefficient:

[0146] θ evap (d,t)=f LightGBM (x evap (d,t))

[0147] Where: θ evap (d, t) is the predicted evaporator fan rapid adjustment coefficient, f LightGBM (·) is the trained LightGBM regressor, is the normalized feature vector.

[0148] Furthermore, in the online adjustment capability evaluation model of the cold storage evaporator fan based on LightGBM proposed in this embodiment, the evaporator fan mainly maintains the air flow in the cold storage. Short-term air volume fluctuations have little effect on temperature control, but local temperature difference accumulation or frost problems need to be avoided. The load adjustment range is wide and can be started and stopped in groups for control.

[0149] The evaporator fan's upward and downward adjustment capacity assessments are calculated as follows:

[0150]

[0151] Where: C up,evap (d, t) is the evaporator fan's upward adjustment capacity, C down,evap (d,t) is the evaporator fan's down-regulation capability, θ evap η is the evaporator fan rapid adjustment coefficient (usually 0.8 to 1.0), reflecting the adjustable ratio after considering the equipment inertia and safety margin; evap (d, t) is the current load rate of the evaporator fan, ζ evapThe evaporator fan rapid reduction coefficient (usually 0.8 to 0.9) reflects the maximum acceptable reduction ratio to avoid temperature control deterioration and P evap,rated is the rated power of the evaporator fan.

[0152] Specifically, for step 4, the AI ​​evaluation model for the online regulation capability of the freezing plant proposed in this embodiment obtains the online regulation capability of the refrigeration compressor unit, the online regulation capability of the condenser fan, and the online regulation capability of the cold storage evaporator fan through the LSTM-based online regulation capability evaluation model of the freezing plant, the XGBoost-based online regulation capability evaluation model of the freezing plant, and the LightGBM-based online regulation capability evaluation model of the freezing plant cold storage evaporator fan. The AI ​​evaluation model is further constructed to evaluate the online regulation capability of the freezing plant:

[0153]

[0154] Where: C up,total (t) is the total power increase, C down,total (t) is the total power reduction, C up,comp (t) is the compressor's adjustment capacity, C up,cond (t) is the condenser's adjustment capacity, C up,evap (t) is the evaporator's adjustment capacity, C down,comp (t) is the compressor's down-regulation capability, C down,cond (t) is the condenser's down regulation capability, C down,evap (t) is the turn-down capacity of the evaporator.

[0155] In an embodiment of the present invention, an electronic device is also provided, including: one or more processors; a storage device on which one or more programs are stored; when the one or more programs are executed by the one or more processors, the one or more processors implement the AI ​​evaluation method for the online load regulation capability of the refrigeration plant described in any of the above embodiments.

[0156] In an embodiment of the present invention, a computer-readable storage medium is further provided, on which a computer program is stored. When the program is executed by a processor, the steps of the AI ​​evaluation method for online load regulation capability of a refrigeration plant in any of the above embodiments are implemented.

[0157] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. An AI evaluation method for online load regulation capability of a refrigeration plant, characterized in that: The steps include: Step 1: Build an LSTM-based model for evaluating the online adjustment capability of refrigeration compressor units in refrigeration plants. This model uses an LSTM-based model to predict the maximum allowable increase in compressor power. This model uses compressor operating data from continuous time windows and inputs it into an LSTM network to extract temporal dynamic features. This model then performs feature compression mapping and predicts the maximum allowable increase in compressor power at the current moment, thereby evaluating the online adjustment capability of the refrigeration compressor. Step 2: Build an XGBoost-based model to evaluate the online regulation capability of condenser fans in refrigeration plants. This model uses an XGBoost-based condenser fan regulation index prediction model to predict the regulation index of the condenser fan load rate and power relationship, and evaluate the online regulation capability of the condenser fan. This model also evaluates the online regulation capability of the condenser fan by capturing the nonlinear response of fan load changes to power changes. Step 3: Construct a LightGBM-based online adjustment capability evaluation model for the evaporator fan in the cold storage of the freezing plant: The LightGBM-based evaporator fan rapid increase coefficient prediction model is used to predict the rapid increase coefficient of the evaporator fan under specific operating conditions. The online adjustment capability of the cold storage evaporator fan is evaluated by comprehensively considering the impact of changes in the cold storage temperature control environment. Step 4. Build an AI evaluation model for the online regulation capability of the refrigeration plant: Based on the evaluated online regulation capability of the refrigeration compressor unit, the online regulation capability of the condenser fan, and the online regulation capability of the cold storage evaporator fan, the online regulation capability of the refrigeration plant is evaluated to provide support for online grid regulation.

2. The AI ​​evaluation method for online load regulation capability of a refrigeration plant according to claim 1 is characterized in that: In step 1, a LSTM-based prediction model for the maximum allowable increase in compressor power is constructed, which includes the following sub-steps: Step 101, feature collection within the time window: collecting compressor operating features for k consecutive time steps, including compressor power, head temperature, refrigerant inlet pressure and outlet pressure, to form an original feature matrix; At time t on day d, the compressor characteristics of the latest k steps are expressed as follows: X comp (d,t)={[P comp (d,τ),T comp (d,τ),P refrig.in (d,τ),P refrig.out (d,τ)]∣τ=t-k+1,...,t} Where: X comp (d, t) is the compressor characteristic matrix, P comp (d,τ),T comp (d, τ) is the power and temperature of the compressor at the time τ on the dth day, P refrig.in (d,τ),P refrig.out (d, τ) are the input pressure and output pressure of the compressor at time τ on day d; Step 102: Normalize the feature matrix. The formula is as follows: X comp (d,t)=Norm(X comp (d,t)) Where: is the normalized compressor characteristic matrix, X j (d,τ) is the compressor characteristic matrix before normalization, μ j is the mean of the j-th class feature, σ j is the standard deviation of the j-th class feature, X comp (d, t) is the characteristic matrix after the compressor is normalized, X comp (d, t) is the characteristic sequence matrix before compressor normalization; Step 103, LSTM dynamic feature encoding: Input the normalized compressor feature matrix into the LSTM network, extract the dynamic evolution characteristics of the system operation through time step recursion, and obtain the hidden state at each moment: Where: h τ is the hidden state vector at the τth moment, c τ is the memory state of the unit at the τth moment, LSTMCell(·) is the LSTM network, X comp (d,τ) is the normalized compressor characteristic matrix, h τ-1 is the hidden state vector at the τ-1th moment, c τ-1 is the unit memory state at the τ-1th moment, h comp (d, t) is the hidden state vector of LSTM at the last moment; Step 104: Feature compression mapping: compress the hidden layer state output by the LSTM into an intermediate prediction vector through a fully connected layer: z comp (d,t)=W z ·h comp (d,t)+b z Where: z comp (d,t) is the intermediate feature compression vector, W z is the feature map weight matrix, b z is the feature map bias vector, h comp (d, t) is the hidden state vector of LSTM at the last moment; Step 105: Prediction of the maximum allowable power increase: Map the intermediate feature vector to a scalar through another fully connected mapping and output it to predict the maximum power increase capability of the compressor: Where: ΔP comp,max (d, t) is the predicted maximum allowable rapid increase in compressor power change, w out is the output weight vector, b out is the output bias term, z comp (d,t) is the intermediate feature compression vector.

3. The AI ​​evaluation method for online load regulation capability of a refrigeration plant according to claim 2 is characterized in that: In step 1, the online adjustment capability of the compressor is evaluated. Based on the dynamic mode of load, temperature, and pressure changes of the compressor in a short period of time, the compressor power and load are adjusted. The rapid adjustment capability is limited by the pressure fluctuation protection of the refrigeration system. It is expressed as follows: Where: C up,comp (d,t),C down,comp (d, t) is the compressor's adjustment capacity and reduction capacity at time t on day d, η comp,max is the maximum allowable load rate of the compressor, η comp (d,t) is the efficiency of the compressor at time t on day d, P comp,rated is the rated power of the compressor, ΔP comp,max is the maximum allowable increase in compressor power, η comp,min The minimum allowable load rate of the compressor.

4. The AI ​​evaluation method for online load regulation capability of a refrigeration plant according to claim 1 is characterized in that: In step 2, a condenser fan upward characteristic index prediction model based on XGBoost is constructed. The specific steps are as follows: Step 201: Raw data collection: Collect the load rate and power data points of the wind turbine over a period of time as the raw data set: Where: D is the original data set, recording the load rate and power; is the sampling time set, η cond (d, t) is the working efficiency of the condenser at time t on day d, P cond (d,t) is the power consumption of the condenser at time t on day d; Step 202, feature transformation: perform logarithmic transformation on the original data set, linearize the power function relationship, and obtain a logarithmic transformed data set: Where: is the working efficiency of the condenser at time t on day d after logarithmic transformation, is the power consumption of the condenser at time t on day d after logarithmic transformation, D is the new data set obtained by logarithmic transformation of the original data set D, and log(·) is the logarithmic function; Step 203: Organize the training features and label sets: Organize the logarithmically transformed data set into feature vectors X and label vectors Y: Step 204: Train the XGBoost regression model based on the feature set: train the XGBoost regression model to learn the mapping relationship between load rate and power, and use the trained regression function f XGB (·) Output: Where: f XGB (·) is the trained XGBoost regression function, ε(d,t) is the fitted residual noise term; Step 205, local derivative calculation: During prediction, the trained regression function is differentiated near the current load rate, the power exponent is extracted, and the upward characteristic index γ of the fan is obtained: Where: γ(d,t) is the condenser fan upward characteristic index at time t on day d, is the local first-order derivative of the regression function.

5. The AI ​​evaluation method for online load regulation capability of a refrigeration plant according to claim 4 is characterized in that: In step 2, the online adjustment capability of the condenser fan is evaluated. The power of the condenser fan is approximately in a square relationship with the wind speed. The upward adjustment capability increases nonlinearly with the remaining wind speed space, and the downward adjustment capability increases linearly with the remaining wind speed space, which can be expressed as: Where: C up,cond (d,t),C down,cond (d, t) is the upward and downward adjustment capabilities of the condenser fan, η cond (d,t) is the original load rate of the condenser fan, P cond,rated is the rated power of the condenser fan, and γ is the fan sharing function adjustment index.

6. The AI ​​evaluation method for online load regulation capability of a refrigeration plant according to claim 1 is characterized in that: In step 3, in the LightGBM-based evaporator fan rapid increase coefficient prediction model, the factors affecting the evaporator fan rapid increase coefficient include: initial load rate, temperature change rate in the cold storage, humidity change rate in the cold storage, and cold storage environment capacity. By collecting the evaporator fan operation characteristics, the evaporator fan rapid increase coefficient is predicted. The specific process is as follows: Step 301, data collection: Collect the evaporator fan operation data at time t on day d, including: the current load rate of the evaporator fan, the temperature change rate after the evaporator fan is adjusted upward, and the humidity change rate after the evaporator fan is adjusted upward, and save the data into the feature vector: x evap (d,t)=[η evap (d,t),ΔT evap (d,t),ΔH evap (d,t)] Where: x evap (d,t) is the eigenvector at a single moment, η evap (d,t) is the current load rate of the evaporator fan, ΔT evap (d,t) is the temperature change rate after the evaporator fan is adjusted upward, ΔH evap (d, t) is the rate of change of humidity after the evaporator fan is adjusted upward; Step 302: Feature normalization: Normalize each feature type. The formula is as follows: Where: is the scalarized eigenvector, x j (d,t) is the eigenvector before scalarization, μ j is the mean of the j-th feature, σ j is the standard deviation of the j-th class feature; Step 303: Organize the training data set: Organize all samples collected over a period of time into a training set and output the training feature set X evap and training label set Y evap : Where: is the normalized eigenvector, θ evap,obs (d, t) is the rapid upward adjustment coefficient obtained from historical experiments or empirical observations; Step 304, LightGBM regression training: Based on the training set, train the LightGBM regressor to learn the mapping relationship from features to up-adjusted coefficients, expressed as: θ evap,obs (d,t)=f LightGBM (x evap (d,t))+ε(d,t) Where: f LightGBM (·) is the trained LightGBM regressor, ε(d,t) is the residual error; Step 305: Predict the evaporator fan rapid increase coefficient: Under the new operating state, input the normalized characteristics and predict the rapid increase coefficient: θ evap (d,t)=f LightGBM (x evap (d,t)) Where: θ evap (d, t) is the predicted evaporator fan rapid adjustment coefficient.

7. The AI ​​evaluation method for online load regulation capability of a refrigeration plant according to claim 6 is characterized in that: In step 3, the online adjustment capacity evaluation model of the evaporator fan in the cold storage of the freezing plant based on LightGBM is used to adjust the load through the start-stop group control, and the upward and downward adjustment capacity of the evaporator fan is evaluated: Where: C up,evap (d,t),C down,evap (d, t) is the evaporator fan's upward and downward adjustment capabilities, θ evap is the evaporator fan rapid adjustment coefficient, ζ evap is the evaporator fan rapid down-regulation coefficient, P evap,rated is the rated power of the evaporator fan.

8. The AI ​​evaluation method for online load regulation capability of a refrigeration plant according to claim 7 is characterized in that: In step 4, an AI evaluation model for the online regulation capability of the refrigeration plant is constructed to evaluate the online regulation capability of the refrigeration plant. The formula is as follows: Where: C up,total (t), C down,total (t) is the overall power increase and power decrease, C up,comp (t), C down,comp (t) is the upward and downward adjustment capabilities of the compressor, C up,cond (t), C down,cond (t) is the upward and downward adjustment capabilities of the condenser, C up,evap (t), C down,evap (t) is the upward and downward adjustment capabilities of the evaporator.

9. An electronic device, characterized in that: include: one or more processors; a storage device having one or more programs stored thereon; When the one or more programs are executed by the one or more processors, the one or more processors implement the AI ​​evaluation method for online load regulation capability of a refrigeration plant as described in any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that A computer program is stored thereon, and when the program is executed by a processor, the steps in the AI ​​evaluation method for online load regulation capability of a refrigeration plant according to any one of claims 1 to 8 are implemented.

Images