Adaptive optimization method for operation and maintenance energy management

Abstract

The invention discloses a self-adaptive optimization method for operation and maintenance energy management, and the method comprises the steps: constructing a hybrid prediction model based on a convolutional neural network and a long-short-term memory neural network through collection and preprocessing of operation data, and improving the prediction accuracy of an energy consumption trend; an optimization control model is generated by using a depth deterministic strategy gradient algorithm, and an optimal control sequence is obtained by solving an optimization problem, so that real-time and self-adaptive adjustment of operation parameters of the energy consumption equipment is realized, and the goal of minimizing energy consumption is achieved. According to the method, the deep learning and reinforcement learning technologies are introduced, the limitation of a traditional energy management method is broken through, accurate prediction and efficient control of energy consumption are achieved, the intelligent level of an energy management system is improved, and the energy consumption and the operation cost are reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of artificial intelligence and energy management, and in particular to an adaptive optimization method for operation and maintenance energy management. Background Art

[0002] With the continuous growth of global energy demand and the enhancement of environmental protection awareness, energy management systems play a vital role in industrial production, building operations and infrastructure maintenance. Traditional energy management mainly relies on preset operating strategies and rules of thumb, lacking the ability to deeply analyze real-time data and dynamically optimize.

[0003] In recent years, with the rapid development of the Internet of Things, big data and artificial intelligence technologies, operation and maintenance energy management systems have gradually evolved towards intelligence and adaptability. Advanced sensor technology enables a large amount of operating data to be collected in real time, and the application of machine learning and deep learning algorithms makes it possible to mine potential patterns from data, predict energy consumption trends and optimize control strategies.

[0004] At present, in terms of optimization control, traditional methods such as model predictive control (MPC) and rule-based control usually rely on precise mathematical models or preset rules and lack the ability to adapt to dynamic changes in the system. Deep reinforcement learning algorithms, such as deep Q-network (DQN) and deep deterministic policy gradient (DDPG), provide new ideas for solving optimization control problems in continuous action spaces, but their application in the field of energy management still faces challenges, such as unstable training process and slow convergence.

[0005] However, existing technologies still have shortcomings in comprehensively utilizing multiple deep learning models to improve prediction accuracy, and in combining advanced reinforcement learning algorithms to achieve real-time, adaptive optimization control. On the one hand, a single prediction model is difficult to fully mine useful information in multi-source heterogeneous data and cannot meet the needs of high-precision prediction; on the other hand, existing optimization control methods often find it difficult to balance computational efficiency and control effects when dealing with high-dimensional, continuous control variables, resulting in energy consumption that cannot reach the optimal level; in addition, the lack of an effective adaptive mechanism to cope with dynamic changes in the system environment and operating conditions affects the overall performance of the energy management system.

[0006] In summary, the existing operation and maintenance energy management technology has limitations in prediction accuracy, adaptive optimization control and system dynamic response. The present invention proposes an adaptive optimization method for operation and maintenance energy management, aiming to solve the above problems. Summary of the invention

[0007] The purpose of this section is to summarize some aspects of the embodiments of the present invention and briefly introduce some preferred embodiments. Some simplifications or omissions may be made in this section and the abstract of the specification and the title of the invention of this application to avoid blurring the purpose of this section, the abstract of the specification and the title of the invention, and such simplifications or omissions cannot be used to limit the scope of the present invention.

[0008] In view of the above existing problems, the present invention is proposed.

[0009] In order to solve the above technical problems, the present invention provides the following technical solutions: collecting operation data in the operation and maintenance energy management system, and performing data preprocessing on the data to form a feature data set;

[0010] Inputting the feature data set into a hybrid prediction model constructed based on a convolutional neural network and a long short-term memory neural network, and outputting a prediction value about the future energy consumption trend;

[0011] Based on the predicted value and the current operating state of the system, a deep deterministic policy gradient algorithm is used to generate an optimization control model, and an optimal control sequence is obtained by solving the optimization problem;

[0012] The optimal control sequence is applied to the operation and maintenance energy management system, with the goal of minimizing energy consumption, and the operating parameters of energy-consuming equipment are adjusted in real time through an adaptive control mechanism.

[0013] As a preferred solution of the adaptive optimization method for operation and maintenance energy management described in the present invention, forming the feature data set includes:

[0014] Collecting operation data from the operation and maintenance energy management system, and performing data cleaning, missing value filling and data normalization processing on the data;

[0015] Construct an initial feature set for the normalized data features;

[0016] The initial feature set is reduced in dimension by principal component analysis, and the main k principal components are selected to form a feature data set after dimension reduction.

[0017] As a preferred solution of the adaptive optimization method for operation and maintenance energy management described in the present invention, the operation data at least includes status data of energy-consuming equipment, energy consumption data, environmental parameters, real-time energy prices, production plans and historical operation and maintenance records.

[0018] As a preferred solution of the adaptive optimization method for operation and maintenance energy management described in the present invention, the constructed hybrid prediction model includes an input layer, a convolutional layer, an LSTM layer and an output layer.

[0019] As a preferred solution of the adaptive optimization method for operation and maintenance energy management described in the present invention, it specifically includes:

[0020] Input a feature data set in the input layer, the shape of which is (n, k), where n is the number of time steps and k is the number of principal components;

[0021] Performing a one-dimensional convolution operation on the input feature data set in the convolution layer to extract local time series features;

[0022] Capturing long-term dependent time series features in the LSTM layer;

[0023] The final hidden state of the LSTM layer is mapped to the predicted value through the output layer.

[0024] As a preferred solution of the adaptive optimization method for operation and maintenance energy management described in the present invention, the calculation formula of the predicted value is:

[0025]

[0026] Among them, W fc is the weight matrix of the fully connected layer, b fc is the bias term, is the predicted value of energy consumption at future time t+Δt, h t is the final hidden state of the LSTM layer.

[0027] As a preferred solution of the adaptive optimization method for operation and maintenance energy management described in the present invention, since energy consumption is a non-negative real number, the predicted value The ratio R of the predicted value to the current energy consumption is:

[0028]

[0029] Among them, if R>1, it means that energy consumption will increase in the future; if R=1, it means that energy consumption will remain unchanged; if R<1, it means that energy consumption will decrease in the future.

[0030] As a preferred solution of the adaptive optimization method for operation and maintenance energy management described in the present invention, the optimal control sequence is applied to the operation and maintenance energy management system, and the operating parameters of the energy-consuming equipment are adjusted in real time through the adaptive control mechanism, including:

[0031] For the current state, the optimal action is calculated through the actor network;

[0032] Map continuous action values ​​to actual control instructions;

[0033] Sending the control instruction to the control system of the device through an industrial communication protocol;

[0034] After receiving the control instruction, the device adjusts the operating parameters.

[0035] As a preferred solution of the adaptive optimization method for operation and maintenance energy management described in the present invention, the operating parameters include the start and stop status, operating speed, load level and working mode of the energy-consuming equipment.

[0036] Beneficial effects of the present invention:

[0037] 1. Eliminate noise and outliers in the data, fill in missing data, and standardize data formats to improve data reliability and availability. The preprocessed feature data set provides a solid foundation for subsequent model training, ensuring that the model can accurately learn the operating rules of the system and improve the performance and robustness of the prediction model;

[0038] 2. Use CNN to extract local patterns and spatial correlations of data, and use LSTM to capture long-term dependencies in time series, so as to make more accurate predictions of future energy consumption trends. By improving prediction accuracy, it is possible to more effectively predict energy demand, optimize resource allocation, and enhance the prediction capability and decision support level of the energy management system;

[0039] 3. Utilize the self-learning ability of the reinforcement learning algorithm to generate a highly adaptable control strategy that can be adjusted in real time to adapt to environmental changes. By solving the optimization problem, the optimal control sequence that minimizes energy consumption is obtained, which provides a scientific basis for actual energy scheduling and equipment control, optimizes system performance and reduces operating costs;

[0040] 4. According to the dynamic changes and optimization strategies of the system, the start and stop, operating speed, load and other parameters of the equipment are controlled in real time, so as to effectively reduce energy waste and improve energy utilization efficiency. Through continuous adaptive optimization, it is ensured that the system can maintain the best operating state under different load and environmental conditions, reduce energy consumption, extend equipment life and improve system reliability. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative labor. Among them:

[0042] Figure 1 It is a flow chart of the adaptive optimization method for operation and maintenance energy management shown in the present invention;

[0043] Figure 2It is a schematic diagram of energy consumption comparison shown in the present invention;

[0044] Figure 3 This is a trend chart of the energy saving rate curve shown in the present invention. DETAILED DESCRIPTION

[0045] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the drawings. It is obvious that the described embodiments are only part of the embodiments of the present invention, but not all of the embodiments.

[0046] Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in the art without making any creative work should fall within the scope of protection of the present invention.

[0047] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0048] According to an embodiment of the present invention, Figure 1 The flowchart shown is an adaptive optimization method for operation and maintenance energy management, which specifically includes the following steps:

[0049] S1. Collect the operation data in the operation and maintenance energy management system, and pre-process the data to form a feature data set. Among them, what needs to be explained in this step is:

[0050] Collecting operation data from the operation and maintenance energy management system:

[0051] Status data of energy-consuming equipment S device (t): including the start and stop status, operating speed, load level, and working mode of the equipment;

[0052] Energy consumption data E(t): actual energy consumption of the equipment at time t;

[0053] Environmental parameters Env(t): such as temperature, humidity, and air pressure environmental conditions;

[0054] Real-time energy prices energy (t): current market price of energy;

[0055] Production Plan(t): the scheduled operation plan of the equipment in the future;

[0056] Historical operation and maintenance records history (t): Past maintenance and failure records;

[0057] Furthermore, statistical methods (such as the triple standard deviation method) are used to detect the collected operating data, and outliers are eliminated, and duplicate records are checked and deleted to ensure the uniqueness of the data;

[0058] Use the mean or median of the feature to fill in the data with missing values. For the i-th sample in feature X:

[0059]

[0060] Among them, X i is the value of the i-th sample in feature X, is the mean of feature X, N valid is the number of samples with non-missing values ​​in feature X, X j is the jth non-missing value sample in feature X;

[0061] For example, suppose feature X has 5 samples, namely [10, 12, missing, 14, 13], and calculate the mean:

[0062]

[0063] Fill missing values: Replace missing values ​​with 12.25,

[0064] Categorical data: Use the mode (the category with the highest frequency) to fill in;

[0065] Scale the numerical features to the interval [0,1] to eliminate the impact of dimensions between different features;

[0066] Furthermore, all preprocessed features are combined into the initial feature set F initial :

[0067] F initial =[S device (t),E(t),Env(t),P energy (t),Plan(t),M history (t)]

[0068] Each sample corresponds to a time point t;

[0069] The initial feature set is organized into a matrix form X with a shape of (n, d), where n is the number of samples, d is the number of features, and each row of X i is the feature vector of the i-th sample;

[0070] Subtract the mean of each feature to make the data mean zero:

[0071]

[0072] in, is the feature mean vector, with shape (1, d),

[0073] Compute the covariance matrix:

[0074]

[0075] Among them, Σ is the covariance matrix, with shape (d,d), For X centered The transposed matrix of

[0076] Perform eigenvalue decomposition on the covariance matrix Σ:

[0077] Σ=QΛQ -1

[0078] Where Λ is a diagonal matrix containing the eigenvalues ​​of the covariance matrix λ1≥λ2≥…≥λ d , Q is the corresponding eigenvector matrix, with a shape of (d, d), Q -1 is the inverse matrix of Q (for orthogonal matrix Q -1 =Q T );

[0079] Calculate the variance contribution rate corresponding to each eigenvalue:

[0080]

[0081] According to the cumulative contribution rate, select the first k eigenvectors Q k , then the cumulative contribution rate is:

[0082]

[0083] When the cumulative contribution rate reaches a set threshold (such as 95%), the value of k is determined;

[0084] X centered Project it to the selected principal component space to obtain the reduced-dimensional feature data set F(t):

[0085] F(t)=X centered Q k

[0086] Among them, Q k are the first k selected feature vectors with a shape of (d, k), and F(t) is the feature data set after dimensionality reduction with a shape of (n, k).

[0087] As an example, suppose there are 100 samples (n=100), each sample has 10 features (d=10), and a matrix X is constructed with a shape of (100,10). X is centered to obtain X centered, calculate the covariance matrix Σ, the shape is (10,10), perform eigenvalue decomposition on Σ, get Λ and Q, select the first k principal components required for the cumulative contribution rate to reach 95%, assuming k = 5, and replace X centered Project it into the new 5-dimensional space and get F(t) with a shape of (100,5).

[0088] S2. Input the feature data set into the hybrid prediction model constructed based on the convolutional neural network and the long short-term memory neural network, and output the predicted value of the future energy consumption trend. It should be noted that the constructed hybrid prediction model includes an input layer, a convolutional layer, an LSTM layer, and an output layer (fully connected layer), where:

[0089] Input Layer:

[0090] Input feature data set F(t), shape is (n,k), where n is the number of time steps and k is the number of principal components;

[0091] Convolutional Layer:

[0092] Perform a one-dimensional convolution operation on the input data to extract local time series features:

[0093] F cnn (t) = σ(W conv *F(t)+b conv )

[0094] Among them, W conv is the convolution kernel weight, and its shape is (f size ,k),f size is the convolution kernel size, b conv is the bias term, σ is the activation function, such as ReLU;

[0095] LSTM Layer:

[0096] The feature sequence output by the CNN part is input into the LSTM layer to capture the time series features with long-term dependencies.

[0097] The calculation of the LSTM layer is as follows:

[0098] i t =σ(W i x t +U i h t-1 +b i )

[0099] f t =σ(W f x t +U f h t-1 +b f )

[0100] o t =σ(W o x t +U o h t-1 +b o )

[0101]

[0102] h t =o t ⊙tanh (c t )

[0103] Among them, x t For F cnn (t) is the current input, h t-1 is the hidden state at the previous moment, c t-1 is the memory unit state at the previous moment, i t 、f t , o t are the activation values ​​of the input gate, forget gate, and output gate, respectively. is the candidate memory cell state, c t 、h t are the current memory unit state and hidden state respectively, W, U, b are weight matrices and bias terms, σ is the Sigmoid activation function, tanh is the hyperbolic tangent activation function, and ⊙ is element-wise multiplication;

[0104] Output layer:

[0105] The last hidden state h of the LSTM layer t Mapped to the predicted value through the fully connected layer:

[0106]

[0107] Among them, W fc is the weight matrix of the fully connected layer, b fc is the bias term, is the predicted value of energy consumption at future time t+Δt;

[0108] Furthermore, the constructed hybrid prediction model needs to be trained, including the following steps:

[0109] (1) The feature data set F(t) and the corresponding actual energy consumption E(t) are combined into training samples and divided into training set and validation set;

[0110] (2) Input F(t) and calculate the predicted value through CNN and LSTM layers

[0111] (3) Use mean square error as the loss function:

[0112]

[0113] Where N is the number of training samples, is the energy consumption of the i-th sample predicted by the model, E i is the actual energy consumption of the i-th sample;

[0114] (4) Calculate the gradient of the loss function with respect to the model parameters;

[0115] (5) Use optimization algorithms (such as Adam) to update model parameters;

[0116] Repeat steps (2) to (5) until the model converges;

[0117] It should also be noted that in this embodiment, for new feature data F(t), the predicted value is calculated using the trained model:

[0118]

[0119] Among them, f CNN-LSTM is the trained hybrid prediction model;

[0120] Since energy consumption is a non-negative real number, the predicted value

[0121] When the ratio R of the predicted value to the current energy consumption is:

[0122]

[0123] Judgment conditions:

[0124] If R>1, it indicates an upward trend in energy consumption in the future;

[0125] If R = 1, it means that energy consumption remains unchanged;

[0126] If R<1, it indicates a downward trend in energy consumption in the future.

[0127] As an example, when R=1.1, future energy consumption increases by 10%; when R=0.9, future energy consumption decreases by 10%.

[0128] S3. Based on the predicted value and the current operating state of the system, the deep deterministic policy gradient algorithm is used to generate an optimization control model, and the optimal control sequence is obtained by solving the optimization problem. Among them, this step also needs to be explained:

[0129] The Deep Deterministic Policy Gradient (DDPG) algorithm is used, which includes the following steps:

[0130] Randomly initialize the actor network policy function μ(s t|θ μ ) and its parameter θ μ ;

[0131] Randomly initialize the critic network value function Q(s t ,a t |θ Q ) and its parameter θ Q ;

[0132] Initialize the target network: actor target network parameters θ μ′ ←θ μ , critic target network parameters θ Q′ ←θ Q ;

[0133] Initialize experience replay pool D;

[0134] For each time step t, the following steps are repeated:

[0135] Select an action based on the current policy and exploration noise:

[0136]

[0137] Among them, a t is the action vector at time t, i.e., the operating parameters of the energy-consuming device, μ(s t |θ μ ) is the actor network based on the state s t The deterministic policy output, θ μ is the set of parameters of the actor network, is the exploration noise at time t;

[0138] For example, the action vector a t Contains operating parameters of energy-consuming equipment: start and stop status a start / stop , running speed a speed , load level a load And working mode a mode ;

[0139] Action a t Applied to energy-consuming equipment to obtain new operating parameters;

[0140] Observe the new state s after executing the action t+1 and instant rewards t ;

[0141] Reward function:

[0142] r t =-(E(t+1)+λ·C(a t ))

[0143] Among them, r tis the instantaneous reward at time t, (E(t+1) is the actual energy consumption at time t+1, λ is the penalty coefficient used to balance the impact of energy consumption and constraint violation, λ≥0, C(a t ) is action a t The penalty term for violating the constraint condition. If the constraint is not violated, then C(a t )=0;

[0144] Will (s t ,a t ,r t ,s t+1 ) is stored in the experience replay pool D;

[0145] Randomly sample a mini-batch of M experiences (s i ,a i ,r i ,s i+1 );

[0146] The critic network is updated to calculate the target value y i :

[0147] y i =r i +γQ′(s i+1 ,μ ′ (s i+1 |θ μ′ )|θ Q′ )

[0148] Among them, y i is the target Q value of the i-th sample, r i is the immediate reward of the i-th sample, γ is the discount factor, 0<γ≤1, and represents the present value of future rewards, Q′(s i+1 ,μ′(s i+1 |θ μ′ )|θ Q′ For the target critic network, the next state s i+1 and the estimated Q-value of the corresponding action, θ μ′ ,θ Q′ are the parameters of the target actor and critic networks;

[0149] Calculate the loss function L:

[0150]

[0151] Among them, L is the average loss function of the critic network, Q(s i ,a i |θ Q ) is the critic network for the sample (s i ,a i )’s estimated Q value, θQ are the parameters of the critic network;

[0152] Use gradient descent to minimize the loss function L and update the critic network parameters;

[0153] Actor network update, calculating policy gradient:

[0154]

[0155] in, is the gradient of the policy, with respect to the actor network parameters θ μ The gradient estimate of is the gradient of the critic network for action a, evaluated at state s i and action a=μ(s i ), θ is the parameter of the actor network μ The gradient of , based on the state s i ;

[0156] Using gradient ascent, update the actor network parameters θ μ , to maximize the cumulative reward;

[0157] Update the target critic network parameters:

[0158] θ Q′ ←τθ Q +(1-τ)θ Q′

[0159] Update the target actor network parameters:

[0160] θ μ′ ←τθ μ +(1-τ)θ μ′

[0161] Among them, θ Q′ ,θ μ′ is the parameter of the target network, τ is the soft update coefficient, such as τ = 0.001;

[0162] Furthermore, after multiple iterations of training, the actor network μ(s t |θ μ ) has learned the optimal strategy for any state s t , optimal control action Directly output by the actor network:

[0163]

[0164] in, is the optimal control action at time t.

[0165] As an example, the current running status is known to be:

[0166] Predicted future energy consumption:

[0167] Energy-consuming equipment status: start / stop status a start / stop =1 (the device is running), running speed a speed =80%, load level a load =70%, working mode a mode =2 (comfort mode);

[0168] Environmental parameters: Temperature Temp(t) = 30°C, Humidity Humidity(t) = 50%;

[0169] Real-time energy prices: P energy (t)=0.8 yuan / kWh;

[0170] Construct the state vector:

[0171] s t =[500,{1,80%,70%,2},{30℃,50%},0.8]

[0172] Randomly initialize the parameters of the actor network and the critic network;

[0173] Initialize the target network parameters;

[0174] Initialize experience replay pool D;

[0175] At time step t, do the following:

[0176] Compute actions via an actor network:

[0177] μ(s t |θ μ )={a start / stop =0.9,a speed =75%,a load =65%,a mode =1.5}

[0178] Add exploration noise:

[0179]

[0180] Get the actual action performed:

[0181]

[0182] Start-stop state: 0.8≥0.5, the equipment keeps running;

[0183] Run speed: set to 77%;

[0184] Load level: set to 64%;

[0185] Working mode: take the integer part and select mode 1 (energy saving mode);

[0186] Adjust equipment operating parameters and apply them to air conditioning systems;

[0187] Observe the new state s t+1 (updated equipment status and environmental parameters);

[0188] Measure actual energy consumption: E(t+1)=480kWh (energy consumption reduced);

[0189] Calculate the immediate reward: Assuming that no constraints are violated, C(a t )=0, penalty coefficient λ=1, then:

[0190] r t =-(480+1×0)=-480

[0191] Will (s t ,a t ,r t ,s t+1 ) is stored in the experience replay pool D;

[0192] Randomly sample M experiences from D, including the most recent one;

[0193] Update the critic network and calculate the target value y i : Assuming γ = 0.99, the Q value estimated by the target network is Q′(s i+1 ,μ′(s i+1 )|θ Q′ )=-470, then:

[0194] y i = -480 + 0.99 × ( -470) = -480 - 465.3 = -945.3

[0195] Calculate the loss function L: Assume that the current estimated Q value of the critic network is Q(s i ,a i |θ Q )=-950, then:

[0196] L = (-945.3-(-950)) 2 =4.7 2 =22.09

[0197] Update the critic network parameters θ Q :Use gradient descent method to minimize L;

[0198] Compute the gradient of the critic network with respect to the action Compute the gradient of the actor network with respect to its parameters μ(s i |θ μ ), update the actor network parameters θ μ , using the gradient ascent method, update the parameters to maximize the cumulative reward;

[0199] Update the target network parameters:

[0200] θ Q′ ←0.001θ Q +0.999θ Q′

[0201] θ μ′ ←0.001θ μ +0.999θ μ′

[0202] Using the updated actor network, the current state s t Compute the optimal action:

[0203]

[0204] Start-stop state: 0.85≥0.5, the equipment continues to run;

[0205] Running speed: set to 70%;

[0206] Load level: set to 60%;

[0207] Working mode: take the integer part and select mode 1 (energy saving mode);

[0208] Adjust equipment operating parameters and apply them to air conditioning systems.

[0209] S4. Apply the optimal control sequence to the operation and maintenance energy management system, with the goal of minimizing energy consumption, and adjust the operating parameters of the energy-consuming equipment in real time through an adaptive control mechanism. It should be noted that the operating parameters in this step include the start and stop status, operating speed, load level and working mode of the energy-consuming equipment;

[0210] For the current state t , calculate the optimal action through the actor network:

[0211]

[0212] Map continuous action values ​​to actual control instructions, for example, start and stop status: If ≥0.5, the device starts, otherwise it stops;

[0213] Operating speed:

[0214]

[0215] Load Level:

[0216]

[0217] Working mode: Select the closest The mode number of

[0218] Send control instructions to the control system of the device through industrial communication protocols (such as MODBUS, PROFINET);

[0219] After receiving the instruction, the device adjusts the operating parameters;

[0220] For example, the current state s t It is known that the actor network can be calculated as follows:

[0221]

[0222] Then the start-stop state: 0.8≥0.5, the equipment should start;

[0223] Running speed: set to 75% of maximum speed;

[0224] Load level: set to 50% of maximum load;

[0225] Working mode: 1.2 Approach mode 1, Select mode 1;

[0226] Send command: send the above parameters to the device to adjust the operating status;

[0227] Furthermore, the feedback data after the acquisition device executes the control instruction is: the new state s t+1 , actual energy consumption E(t+1) and environmental change Env(t+1);

[0228] According to the actual energy consumption and constraints, the reward r is calculated t :

[0229]

[0230] in, is the penalty for violating the constraint. If the constraint is not violated, then

[0231] New experience Deposit into experience replay pool D;

[0232] Adopt the same training steps as in step S3 to update the actor and critic networks;

[0233] Enter the next time step t+1 and repeat step S4.

[0234] As an example, suppose that after the action is executed, the actual energy consumption E(t+1) = 100 kWh and the target energy consumption is 95 kWh. If the constraints are not violated, but:

[0235] r t =-100

[0236] Use new experience data to update strategies and improve the quality of next decision.

[0237] It should also be noted that the conventional method adopts an energy management strategy based on fixed parameters and preset rules, that is, it controls according to the established equipment operation plan and parameters, does not consider real-time operation data and environmental changes, and cannot adjust the control strategy according to real-time data and environmental changes, resulting in energy waste or low equipment operation efficiency; does not use advanced prediction models, and is difficult to accurately predict future energy consumption trends and cannot be optimized in advance; and relies on manually set rules, and cannot achieve global optimal control of complex systems;

[0238] In order to verify the effectiveness of an adaptive optimization method for operation and maintenance energy management proposed in an embodiment of the present invention, the embodiment of the present invention selects an energy management system of a large manufacturing plant as an experimental object, and conducts comparative tests on the method of the present invention and the traditional energy management method based on fixed parameters and preset rules;

[0239]

Experimental environment conditions

[0240] Selected objects: Energy management system of a large manufacturing enterprise, including key energy-consuming equipment such as air-conditioning system, production equipment and lighting system;

[0241] Air conditioning system: affects environmental comfort, has high energy consumption, and is greatly affected by ambient temperature;

[0242] Production equipment: directly involved in the production process and has a direct impact on production efficiency;

[0243] Lighting system: ensures the lighting needs of the production area, and the energy consumption is relatively stable;

[0244] Time frame: 7 consecutive days of actual production cycle, covering weekdays and weekends;

[0245] Temperature range: 27℃ to 31℃, simulating different seasons or weather conditions;

[0246] Humidity range: between 40% and 60%, affecting air conditioning load;

[0247] Production plan fluctuations: daily production tasks vary slightly, simulating changes in actual production;

[0248] Equipment running time: dynamically adjusted according to production plan and environmental conditions;

[0249]

Experimental process

[0250] Traditional method implementation process:

[0251] Set fixed equipment operating parameters (such as operating time, power setting);

[0252] Run energy-consuming equipment according to preset parameters without making real-time adjustments;

[0253] Record equipment operation data and energy consumption daily;

[0254] Implementation process of the method of the present invention:

[0255] Collect status data, energy consumption data, environmental parameters, production plans and historical operation and maintenance records of energy-consuming equipment, clean and normalize the data, and form a feature data set;

[0256] Input the feature data set into a hybrid prediction model based on convolutional neural network and long short-term memory neural network, train the model, and output the predicted value of energy consumption in the next hour;

[0257] Based on the predicted value and the current operating status, the deep deterministic policy gradient algorithm is used to generate an optimization control model, solve the optimization problem, and obtain the optimal control sequence;

[0258] Apply the optimal control sequence to the energy management system, adjust the operating parameters of energy-consuming equipment in real time, and record the adjusted equipment operating data and energy consumption on a daily basis;

[0259]

Experimental indicators

[0260] In order to verify the effectiveness of the method of the present invention in reducing energy consumption, improving energy utilization efficiency and adapting to environmental changes, comparisons were made through experimental indicators (energy consumption, energy saving rate);

[0261] Energy consumption (kWh): daily and cumulative energy consumption;

[0262] Energy saving rate (%):

[0263]

[0264] Wherein, ξ is the energy saving rate, Z is the energy consumption of the traditional method, and B is the energy consumption of the method of the present invention;

[0265]

Experimental results

[0266] Reference Figure 2It can be seen intuitively that the energy consumption of the traditional method reaches a peak (1200kWh) on the first day, drops to a minimum (300kWh) on the fourth day, and then shows an upward trend; the energy consumption of the method of the present invention is lower than that of the traditional method every day, and the curve trend is basically consistent with the traditional method, but the overall level is lower;

[0267] It can be seen from the curves that the energy consumption of both the traditional method and the method of the present invention fluctuates with the change of the number of days. The difference between the two curves shows that the method of the present invention achieves energy saving in all time periods, especially on the 4th and 7th days, the energy saving effect is more significant;

[0268] Reference Figure 3 , from the 1st to the 3rd day, the energy saving rate increased steadily, indicating that the energy saving effect of the method of the present invention is gradually manifested. The energy saving rate on the 4th day reached 10%, which may be related to the environmental conditions or production load of the day. The energy saving rate from the 5th to the 7th day remained at a high level, indicating that the method of the present invention is continuously adapting to system changes and improving the energy saving effect;

[0269] Preferably, the method of the present invention saves an average of about 110 kWh per day, and a total of 770 kWh. Under different environmental conditions and production loads, the method of the present invention achieves a reduction in energy consumption, and the energy saving rate is increased from 8.33% to 10.13%, which reflects the adaptive optimization capability of the method of the present invention.

[0270] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. An adaptive optimization method for operation and maintenance energy management, characterized in that: include: Collect the operation data in the operation and maintenance energy management system, and perform data preprocessing on it to form a feature data set; Inputting the feature data set into a hybrid prediction model constructed based on a convolutional neural network and a long short-term memory neural network, and outputting a prediction value about future energy consumption trends; Based on the predicted value and the current operating state of the system, a deep deterministic policy gradient algorithm is used to generate an optimization control model, and an optimal control sequence is obtained by solving the optimization problem; The optimal control sequence is applied to the operation and maintenance energy management system, with the goal of minimizing energy consumption, and the operating parameters of energy-consuming equipment are adjusted in real time through an adaptive control mechanism.

2. The adaptive optimization method for operation and maintenance energy management according to claim 1, characterized in that: Forming the feature data set includes: Collecting operation data from the operation and maintenance energy management system, and performing data cleaning, missing value filling and data normalization processing on the data; Construct an initial feature set for the normalized data features; The initial feature set is reduced in dimension by principal component analysis, and the main k principal components are selected to form a feature data set after dimension reduction.

3. The adaptive optimization method for operation and maintenance energy management according to claim 2, characterized in that: The operation data at least includes status data of energy-consuming equipment, energy consumption data, environmental parameters, real-time energy prices, production plans and historical operation and maintenance records.

4. The adaptive optimization method for operation and maintenance energy management according to claim 1, characterized in that: The constructed hybrid prediction model includes an input layer, a convolutional layer, an LSTM layer and an output layer.

5. The adaptive optimization method for operation and maintenance energy management according to claim 4, characterized in that: Specifically include: Input a feature data set in the input layer, the shape of which is (n, k), where n is the number of time steps and k is the number of principal components; Performing a one-dimensional convolution operation on the input feature data set in the convolution layer to extract local time series features; Capturing long-term dependent time series features in the LSTM layer; The final hidden state of the LSTM layer is mapped to the predicted value through the output layer.

6. The adaptive optimization method for operation and maintenance energy management according to claim 5, characterized in that: The calculation formula of the predicted value is: Among them, W fc is the weight matrix of the fully connected layer, b fc is the bias term, is the predicted value of energy consumption at future time t+Δt, h t is the final hidden state of the LSTM layer.

7. The adaptive optimization method for operation and maintenance energy management according to claim 6, characterized in that: Since energy consumption is a non-negative real number, the predicted value The ratio R of the predicted value to the current energy consumption is: Among them, if R>1, it means that energy consumption will increase in the future; if R=1, it means that energy consumption will remain unchanged; if R<1, it means that energy consumption will decrease in the future.

8. The adaptive optimization method for operation and maintenance energy management according to claim 1, characterized in that: The optimal control sequence is applied to the operation and maintenance energy management system, and the operating parameters of the energy-consuming equipment are adjusted in real time through an adaptive control mechanism, including: For the current state, the optimal action is calculated through the actor network; Map continuous action values ​​to actual control instructions; Sending the control instruction to the control system of the device through an industrial communication protocol; After receiving the control instruction, the device adjusts the operating parameters.

9. The adaptive optimization method for operation and maintenance energy management according to claim 1 or 8, characterized in that: The operating parameters include the start / stop status, operating speed, load level and working mode of the energy-consuming equipment.

Images