Air conditioner load prediction and energy saving method based on machine learning

Abstract

The invention relates to the technical field of air conditioner load prediction, and discloses an air conditioner load prediction and energy saving method based on machine learning, and the method comprises the following steps: collecting and preprocessing the historical load, multi-dimensional meteorological data and equipment operation parameters of a building air conditioner; historical loads and meteorological data are input into a GBDT model, meteorological feature importance is calculated, and key features are screened; taking the key meteorological characteristics and the historical load as input, and performing load prediction by using an LSTM model; based on a particle swarm optimization algorithm, operating parameters of the air conditioning system are dynamically optimized; and updating the GBDT and LSTM model when the prediction error exceeds a threshold value or the period arrives. According to the method, key meteorological characteristics are screened through GBDT to improve the load prediction precision, adaptive modeling of different climate areas is achieved in combination with LSTM, the air conditioner operation parameters are dynamically adjusted through the particle swarm optimization algorithm, the comprehensive energy consumption of the air conditioner system is effectively reduced, and the overall energy-saving efficiency and the intelligent level of the air conditioner system are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of air - conditioning load prediction, and particularly to an air - conditioning load prediction and energy - saving method based on machine learning. Background Technique

[0002] With the acceleration of the urbanization process and the continuous expansion of building scale, building energy consumption is increasing day by day. Among them, as the main component of building energy consumption, the energy consumption of air - conditioning systems continues to rise. How to achieve the efficient operation of air - conditioning systems has become a key issue in improving building energy utilization efficiency, reducing operating costs, and reducing carbon emissions.

[0003] At present, traditional air - conditioning load prediction methods are mostly based on empirical rules or static mathematical models, and it is difficult to cope with the high volatility and complexity of load demand in the modern building operation environment. These methods generally have problems such as poor model adaptability and low prediction accuracy, and it is difficult to meet the actual needs of refined energy - saving regulation. At the same time, existing energy - saving control strategies are mostly driven by preset rules, lacking a response mechanism to prediction information, and it is difficult to achieve real - time optimal adjustment of key operating parameters such as the water outlet temperature of the unit and the water pump flow rate, resulting in energy waste and low operating efficiency.

[0004] Air - conditioning load prediction technology can provide a basis for feed - forward regulation of the system by predicting the load change trend in advance, effectively overcoming the lag of traditional feedback control in time response, not only improving the thermal comfort of the indoor environment, but also providing support for energy - saving control. However, there are still significant deficiencies in existing research in terms of feature selection, model generalization ability, and energy - saving linkage control.

[0005] For example, Patent CN114118571A proposes a heating load prediction method based on LSTM. Although LSTM is introduced to improve the prediction accuracy, this method does not fully consider the comprehensive influence of meteorological factors in different regions on building air - conditioning loads, and lacks linkage with energy - saving control algorithms. Summary of the Invention

[0006] The object of the present invention is to provide an air - conditioning load prediction and energy - saving method based on machine learning, which solves the problems of low load prediction accuracy, lack of regional adaptability, and lag in energy - saving control response in the prior art.

[0007] To achieve the above object, the present invention is realized through the following technical solutions: An air - conditioning load prediction and energy - saving method based on machine learning, comprising the following steps: S1. Multi - source data collection, where the multi - source data collection includes collecting historical load data of building air - conditioners, multi - dimensional meteorological data, and equipment operation parameters, and pre - processing the multi - source data; S2. Input the historical load data and meteorological data into the gradient boosting decision tree model, calculate the importance scores of each meteorological feature, and screen out the key meteorological features whose cumulative score ratio reaches a preset threshold; S3. Use the screened key meteorological features and historical load data as inputs, construct a load prediction model using a long short-term memory network, and output the predicted value of the air-conditioning load for the future period; S4. Dynamically optimize the operating parameters of the air-conditioning system based on the particle swarm optimization algorithm; S5. Update the gradient boosting decision tree model and the long short-term memory network model periodically or when the prediction error exceeds a set threshold.

[0008] Preferably, in step S1, the multi-dimensional meteorological data includes temperature, humidity, wind speed, and sunshine intensity; the equipment operating parameters include the inlet water temperature of the unit, the outlet water temperature of the unit, the pump flow rate, and the number of operating units.

[0009] Preferably, in step S1, the data preprocessing includes: Perform data cleaning on the collected multi-source data to eliminate duplicate data and invalid data; Detect and delete outliers based on the isolation forest algorithm; Fill in the missing data using the time series interpolation method, where the time series interpolation method is the cubic spline interpolation method.

[0010] Preferably, in step S2, the preset threshold is 80%.

[0011] Preferably, in step S3, use the historical load data of the previous 24 hours sampled in hours and the screened key meteorological features as inputs, output the predicted value of the air-conditioning load for the next 1 hour, and automatically adjust the hyperparameters of the long short-term memory network using an optimization algorithm.

[0012] Preferably, the hyperparameters include the number of hidden layers, the learning rate, and the Dropout rate, and the optimization algorithm is the whale optimization algorithm.

[0013] Preferably, step S4 includes: Use the predicted load value as the input to construct an objective function to minimize the total energy consumption of the air-conditioning system; Under the constraint condition that the load demand satisfaction rate ≥ 95%, search for the Pareto optimal solution set through the particle swarm optimization algorithm; Dynamically optimize the operating parameters including the number of operating units of the unit, the number of operating units of the pump, the set value of the outlet water temperature, and the operating frequency; And output the energy-saving operation plan that meets the optimization goal in real time.

[0014] Preferably, step S5 includes: Periodically retrain the gradient boosting decision tree model and the long short-term memory network model to adapt to meteorological conditions or seasonal changes; An update is triggered when the prediction error exceeds a set threshold, which is 10%.

[0015] The present invention also provides an air conditioning load prediction and energy-saving system based on machine learning, comprising: Multi-source data acquisition and preprocessing module, used to collect historical building air conditioning load data, multi-dimensional meteorological data and equipment operating parameters, and preprocess the data; A feature selection module is used to input the historical load data and meteorological data into a gradient boosting decision tree model, calculate the importance score of each meteorological feature, and select key meteorological features whose cumulative score reaches a preset threshold; The load forecasting module is used to take the filtered key meteorological characteristics and historical load data as input, build a load forecasting model using a long-short-term memory network, and output the predicted air conditioning load value for the future period; Energy-saving optimization module, which is used to dynamically optimize the operating parameters of the air-conditioning system based on the particle swarm optimization algorithm; The model updating module is used to update the gradient boosting decision tree model and the long short-term memory network model periodically or when the prediction error exceeds a set threshold.

[0016] In summary, the present invention includes at least one of the following beneficial technical effects: 1. This invention utilizes a gradient boosting decision tree-based approach to screen key meteorological features and combine it with a long-short-term memory network modeling solution, significantly improving the accuracy of air conditioning load forecasts under complex climate conditions. Compared to existing approaches that employ inadequate feature selection and employ a single prediction model, this approach addresses the technical limitations of large prediction errors and weak model generalization.

[0017] 2. This invention achieves the technical effect of adaptively constructing a prediction model based on different climate zones by integrating multi-source data with a differentiated screening mechanism based on regional meteorological characteristics, combined with LSTM model training. Compared to existing solutions that lack targeted processing, this solves the model's poor generalizability and adaptability in hot and humid regions and cold and dry regions.

[0018] 3. This invention utilizes a particle swarm optimization algorithm based on load forecasting to dynamically control operating parameters, achieving the technical effect of automatically optimizing the air conditioning system's operating state and reducing energy consumption. Compared to traditional fixed strategies or static parameter settings, this approach addresses the technical issues of extensive energy-saving control and low energy efficiency management in air conditioning systems.

[0019] 4. The present invention introduces a model update mechanism based on error threshold and periodicity, achieving the technical effect of maintaining the long-term effectiveness of the prediction model and adapting to external changes. Compared with the prior art solutions where the model is not updated for a long time and the prediction ability decays, it solves the deficiencies of the decreasing load prediction accuracy over time and the poor system response ability. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 It is a schematic flowchart of the method of the present invention; Figure 2 It is a schematic diagram of an air-conditioning load prediction and energy-saving system based on machine learning of the present invention; Figure 3 It is the overall flowchart of the present invention; Figure 4 It is the system architecture diagram of the present invention; Figure 5 It is an example diagram of the GBDT feature importance ranking of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0021] The following will further describe the present invention in detail Figure 1 - with reference to the Figure 5 drawings.

[0022] The embodiment of the present invention provides a machine learning-based air-conditioning load prediction and energy-saving method, including the following steps: S1. Multi-source data collection, where the multi-source data collection includes collecting historical load data of building air conditioners, multi-dimensional meteorological data, and equipment operation parameters, and preprocessing the multi-source data; S2. Input the historical load data and meteorological data into the gradient boosting decision tree model, calculate the importance scores of each meteorological feature, and screen the key meteorological features whose cumulative score ratio reaches a preset threshold; S3. Use the screened key meteorological features and historical load data as inputs, and construct a load prediction model using a long short-term memory network to output the predicted value of the air-conditioning load in the future period; S4. Dynamically optimize the operation parameters of the air-conditioning system based on the particle swarm optimization algorithm; S5. Update the gradient boosting decision tree model and the long short-term memory network model periodically or when the prediction error exceeds a set threshold.

[0023] For step S1, in this embodiment, step S1 is used to realize the collection and standardization processing of multi-dimensional feature data required for building air-conditioning load prediction, ensuring the integrity, continuity, and availability of the input data, which is the basic link for subsequent model training and optimal control. This step mainly includes operations such as multi-source data collection, data cleaning, outlier detection and removal, missing value filling, and time alignment.

[0024] In this embodiment, the multi-source data collection covers the following three types of data sources: historical load data of the building air-conditioning system, multi-dimensional meteorological data, and air-conditioning system equipment operating parameter data.

[0025] The historical load data refers to the actual cooling load value of the air-conditioning system within a unit time (preferably at the hourly level), which is usually measured and recorded by a building energy management system (BEMS) or a heat meter to characterize the temporal variation characteristics of the building's energy consumption intensity.

[0026] The multi-dimensional meteorological data includes but is not limited to the following variables: Temperature (unit: °C) refers to the outdoor ambient temperature of the building area during the corresponding time period; Humidity (unit: %) is the relative humidity in the air; Wind speed (unit: m / s), which indicates the impact of the external wind environment on building heat exchange; Sunshine intensity (unit: W / m 2 ), which is used to characterize the impact of solar radiation on building loads.

[0027] This type of data can be collected and obtained in real time by accessing the regional weather station API interface or local meteorological sensors, and the data frequency is consistent with the load data.

[0028] The equipment operating parameters include the following information The unit water inlet temperature and unit water outlet temperature (unit: °C) are used to reflect the working conditions of the air conditioning system's hot and cold conversion; Pump flow (unit: m 3 / h), reflecting the system's water supply and return capacity; The number of operating equipment is used to indicate the number of chillers and water pumps currently in operation, and is one of the important variables for subsequent optimized scheduling.

[0029] To ensure time consistency between different data sources, the acquisition system in this embodiment merges and aligns all data according to a unified timestamp, preferably using an hourly alignment mechanism (such as every hour) for unified sampling.

[0030] After data collection is completed, the preprocessing stage begins. The preprocessing process includes the following key steps: First, perform data cleaning. This process is used to eliminate redundant or invalid data, and mainly includes the identification and deletion of the following types of data: Completely duplicate records, that is, multiple identical data items with the same timestamp; Data rows containing significant error markers (such as -9999, null values) or where all fields are zero; Data with missing timestamps or non-standard formats appears.

[0031] After data cleaning, detect and remove possible outliers in the data. In this embodiment, the Isolation Forest algorithm is used to identify outliers.

[0032] Isolation Forest is an unsupervised outlier detection method based on tree structure, which can automatically identify sample points deviating from the overall distribution in the high-dimensional feature space. Its basic principle is: in the constructed random split tree, the average path length required for outliers is significantly shorter than that for normal points.

[0033] The specific algorithm process is as follows: Construct multiple sub-sample random split trees (IsolationTrees), and each tree continuously divides the sample set by randomly selecting features and setting split thresholds; Calculate the average path length of the sample point reaching the leaf node in all trees; The outlier degree is defined as: where s(x,n) represents the outlier score of sample point x; E(h(x)) is its average path length; c(n) is the expected path length of normal samples in a completely random tree; and n is the number of samples.

[0034] Set an empirical threshold, and samples with outlier scores higher than this threshold are determined to be outliers and removed.

[0035] After outlier processing, perform missing value filling. For short-term missing samples in the time series, the Cubic Spline Interpolation method is used for interpolation.

[0036] This method constructs a smooth cubic polynomial curve between known points and ensures that the interpolation curve is continuous in the first and second derivatives at each node, which is beneficial to maintaining the time smoothness and trend characteristics of the data.

[0037] The specific settings are as follows: Given the known time series data points {(t1,y1),(t2,y2),...,(t n ,y n )}, the interpolation function S(t) is defined on each sub-interval [t i ,t i+1 as: S i (t)=a i +b i (t-t i )+c i (t-ti ) 2 +d i (t - t i ) 3 ; Among them, the coefficients a i , b i , c i , d i are obtained by solving a tridiagonal linear system of equations, so that the function is continuous at the nodes and its first and second derivatives.

[0038] This interpolation method is applicable to the completion processing of continuous-valued variables and has good smoothness and mathematical interpretability.

[0039] Finally, the processed data above is subjected to unified normalization and time standardization processing. All feature variables are uniformly transformed to the same time scale, and a linear normalization method (such as min-max normalization) is adopted to make the numerical distributions of each variable unified, avoiding the impact on subsequent model training due to dimensional differences.

[0040] To sum up, in this embodiment, multi-source information of the building air-conditioning system is systematically collected, and data cleaning, anomaly processing, and completion are completed relying on algorithms such as Isolation Forest and cubic spline interpolation.

[0041] For step S2, in this embodiment, after completing the multi-source data collection and preprocessing, to improve the training efficiency and generalization ability of the load prediction model, the system inputs the historical load data and meteorological data into the Gradient Boosting Decision Tree (GBDT) model together, and screens out the key meteorological features that have a greater impact on the load change by calculating the importance scores of each meteorological feature in the prediction target.

[0042] In this embodiment, the historical load data is the cold load time series of the air-conditioning system obtained in step S1, denoted as: y = {y1, y2, …, y n}; Among them, y i represents the actual load value corresponding to the i-th moment.

[0043] The meteorological feature data consists of variables such as temperature, humidity, wind speed, and sunshine intensity, and after preprocessing, it constitutes the input feature matrix: Among them, n represents the number of samples, that is, the number of time steps of continuous sampling by the system (preferably sampled hourly); m represents the number of types of meteorological features, for example, including variables such as temperature, humidity, wind speed, and sunshine intensity; x ijDenote the observed value of the \(i\)-th sample in the \(j\)-th meteorological dimension.

[0044] To evaluate the importance of each meteorological feature, in this embodiment, a GBDT regression model is constructed, with the prediction target being \(y\) and the input features being \(X\). The GBDT model is a set of cascaded base learners (weak learners), and each base learner is a regression tree used to fit the residuals of the previous-round model. The overall predicted value of this model is: where \(f\) t (·) is the \(t\)-th round of base learner; \(T\) is the total number of iteration rounds; \(x\) i is the input vector of the \(i\)-th sample.

[0045] where represents the total loss value of the current model on all training samples; \(y\) i represents the true air-conditioning load value at the \(i\)-th time point, which is sourced from historical observed data; represents the predicted load value of the model on the \(i\)-th sample, that is, the output result obtained through the current GBDT model; \(n\) represents the total number of training samples, corresponding to the length of the historical data sequence.

[0046] During the construction of each regression tree, the node splitting is based on the loss decrease value caused by the feature splitting, that is, the "gain". In this embodiment, the cumulative value of this gain is used as an indicator to measure the feature importance.

[0047] Let the \(j\)-th meteorological feature \(x\) j be the set of nodes participating in the split in the \(t\)-th tree, then its corresponding total importance score is defined as: where is the decrease in the loss function brought about by the split of node \(s\), reflecting the contribution of this feature to reducing the fitting error in the current round of the model.

[0048] To facilitate the comparison between features, the importance scores \(I\) of all features j are normalized to obtain the feature proportion score \(W\) j : where \(W\) j represents the importance proportion of the \(j\)-th meteorological feature among all features. After the calculation, all features are sorted in descending order according to \(W\) j , and the weights are accumulated in sequence, and the features whose cumulative scores reach the preset threshold are selected as the key meteorological features.

[0049] In this embodiment, preferably, the preset threshold is set to 80%. That is, starting from the feature with the highest ranking, when the cumulative weight sum first exceeds 80%, the first r features are selected as the key feature set: such that where represents the k-th selected meteorological feature, and the index j k ∈{1,2,…,m}; F key represents the selected set of key meteorological features; is the importance score of the k-th feature sorted by importance.

[0050] To further improve the stability and generalization ability of the feature selection results, this embodiment uses a cross-validation mechanism to evaluate and train the GBDT model. Specifically, the input data is divided into several folds (for example, preferably 5 folds). In each fold, a GBDT model is constructed and the feature scores are calculated. Finally, the average value of each feature score in each fold is taken as the final sorting basis to mitigate the impact of accidental data fluctuations on the feature evaluation results.

[0051] It should be noted that the GBDT model in this embodiment is only used as a feature selection tool and is not used for direct load prediction. Its output results will be used as the input variable set of the subsequent Long Short-Term Memory (LSTM) model to reduce the model input dimension, improve the training efficiency, and improve the prediction stability.

[0052] In addition, compared with unsupervised dimensionality reduction methods such as principal component analysis (PCA), the feature selection based on GBDT retains the physical meaning and interpretability of meteorological features and is more suitable for the interpretable modeling tasks of subsequent energy scheduling control.

[0053] In summary, this embodiment completes the screening of key meteorological variables by constructing a gradient boosting regression model and calculating feature scores, and combines the set cumulative proportion threshold to form a structurally reasonable, logically clear, and data-driven automatic feature screening mechanism.

[0054] For step S3, in this embodiment, after completing the screening of key meteorological features in step S2, a Long Short-Term Memory (LSTM) model is further constructed to predict the future air-conditioning cooling load. This model combines historical load data and meteorological feature sequences and can characterize the temporal sequence and meteorological driving characteristics of load changes.

[0055] The model samples in hours. Specifically, the input at each moment is composed of the following two parts: Historical load observations for the past 24 consecutive hours; Key meteorological characteristic values synchronized with the above time period.

[0056] Let the current prediction time be \(t\), then the historical load sequence can be expressed as: Y hi s t ={y t-24 ,y t-23 ,…,y t-1}; where \(y t-k represents the measured value of the cooling load in the \(k\)th hour before the prediction time, with the unit of kilowatt (kW), and there are 24 dimensions in total. This sequence is used to capture the internal pattern of the system load evolving over time.

[0057] Meanwhile, the input features include the set of key meteorological features screened out in step S2 The construction method, variable definition, and screening logic of the said set are detailed in the aforementioned step S2, and will not be elaborated here.

[0058] Each dimension of the key meteorological features also constitutes a time series of the past 24 hours and is aligned with the historical load by time point. Thus, the input vector for each time step is constructed as: where \(y t-k represents the actual air-conditioning load value in the \(k\)th hour before the prediction time \(t\), with the unit of kilowatt; represents the observed value of the \(j q th key meteorological feature at time \(t - k\); \(r\) is the number of key meteorological features screened out, and its source and screening method have been given in step S2; the vector dimension is \(1 + r\), where 1 dimension is the load value and the remaining \(r\) dimensions are meteorological features; Thus, a complete input sequence tensor is formed: <l where the input vector structure for each time step is \(x t-k , containing one historical load value and \(r\) key meteorological features; \(X in is input into the LSTM network as a whole for modeling the time-dependent relationship in the sequence.

[0059] This tensor is fed into the LSTM model for sequence modeling. The LSTM network is composed of multiple memory units connected in series, and each unit contains a forget gate, an input gate, and an output gate for controlling the retention and update of information. The core update mechanism of LSTM is as follows: f k =\(\sigma(W f [h k-1 ,x k +b f); i k = σ(W i [h k-1 , x k + b i ); o k = σ(W o [h k-1 , x k + b o ); h k = o k ⊙ tanh(c k ); Among them, is the input vector at the k-th time step, which contains the historical load value and r key meteorological features at the corresponding moment; is the hidden state (i.e., the output state) of the previous time step; is the forget gate vector, which is used to control whether the memory in the previous moment state c k-1 is retained; is the input gate vector, which is used to control the degree of writing of new information at the current time step; is the candidate memory vector, which represents the potential new information at the current moment; is the current memory cell state, which combines the previous state and the current input; is the output gate vector, which controls the explicit information of the final output; is the current hidden state, which is the final output of the current time step and also serves as the input for the next step; are the weight matrices corresponding to each gating unit; are the bias terms of each gating unit; σ(·) is the Sigmoid function, and its mapping range is (0, 1), which is used for gating control; tanh(·) is the hyperbolic tangent function, and its output range is (-1, 1), which is used for state non-linear compression.

[0060] After 24 time steps of input, the final hidden state h 24 is sent to the fully connected layer to generate the prediction result: Among them, represents the predicted value of the cooling load at the prediction time t; W y and b y are the weights and biases of the output layer.

[0061] To improve the model performance and reduce the burden of manual hyperparameter tuning, the Whale Optimization Algorithm (WOA) is introduced in this embodiment to automatically optimize the hyperparameters of the LSTM network. The optimized hyperparameters include, but are not limited to: The number of LSTM hidden layer units (i.e., the dimension of h k ); The learning rate, which is used to control the step size of gradient descent; The Dropout rate, which is used for regularization during training to prevent overfitting.

[0062] The Whale Optimization Algorithm simulates the behavior of humpback whales hunting prey, iteratively updates the individual positions in the search space, and realizes the search for the global optimal parameter combination. The objective function of the algorithm is defined as the mean squared error on the validation set: where, y i represents the true load of the i-th sample in the validation set; is the corresponding model prediction value, and N is the number of validation samples.

[0063] The WOA algorithm evaluates the corresponding to the current individual hyperparameter combination in each round and guides the search direction accordingly. Finally, a set of optimal hyperparameters is output for the final training and deployment of the LSTM network.

[0064] After the model training is completed, during actual operation, only the load data of the continuous 24 hours before any moment and the key meteorological feature sequence need to be input into the above network, and the air-conditioning load prediction result for the next 1 hour at the current moment can be obtained.

[0065] For step S4, in this embodiment, a method based on the Particle Swarm Optimization Algorithm (PSO) is proposed to dynamically optimize the operating parameters of the air-conditioning system. This method uses the load prediction value as the input, and searches for the optimal operating parameters of the air-conditioning system through the Particle Swarm Optimization Algorithm to minimize the total energy consumption of the system while ensuring that the load demand satisfaction rate is not less than 95%. During the optimization process, key parameters such as the number of operating air-conditioning units, the number of operating water pumps, the set value of the outlet water temperature, and the operating frequency are adjusted to achieve energy-saving operation of the air-conditioning system.

[0066] In this embodiment, the goal of the Particle Swarm Optimization Algorithm is to minimize the total energy consumption of the air-conditioning system. The following constraint conditions need to be followed during the optimization process: Load demand satisfaction rate: The load demand satisfaction rate of the system must be greater than or equal to 95%. This constraint ensures that the air-conditioning system can fully meet the preset load demand and avoids insufficient load demand caused by the system operating too energy-efficiently. The load demand satisfaction rate is calculated by the following formula: Among them, represents the load prediction value at the i-th moment; y i represents the actual load demand at this moment; n is the prediction duration.

[0067] Energy consumption minimization: The optimization goal is to minimize the total energy consumption of the air conditioning system. The energy consumption of the air conditioning system is related to multiple factors, such as the number of air conditioning units, the number of operating water pumps, the set value of the outlet water temperature, and the operating frequency, etc. The overall energy consumption E total is calculated by the formula: Among them, P unit (k) is the power consumption of the k-th operating unit; t k is the operating time of this unit; N is the total number of operating units.

[0068] The particle swarm optimization algorithm performs global search by simulating swarm behavior to find the optimal solution of the system operating parameters. In this embodiment, each particle in the particle swarm represents a candidate solution, and the dimensions of the solution include control parameters such as the number of operating units of the air conditioning system, the number of operating water pumps, the set value of the outlet water temperature, and the operating frequency. The algorithm flow is as follows: Particle initialization: Randomly initialize the position and velocity of the particles in the search space. The position of the particle represents a possible combination of control parameters, and the velocity represents the direction and amplitude of the particle moving in the solution space. The initial position vector of the particle is: pi = {pi1, pi2,..., pid}; Among them, p id is the position of the i-th particle in the d-th dimension, representing a control parameter of the air conditioning system.

[0069] Velocity update: The velocity update formula of the particle is as follows: Among them, v id is the velocity of particle i in dimension d; w is the inertia weight; c1 and c2 are acceleration constants; rand1, rand2 are random numbers; is the personal historical optimal position of the particle; is the global optimal position.

[0070] Position update: The update formula of the particle is: p id = p id + v id ; By updating the position and velocity of the particles, the particles continuously explore the optimal solution in the solution space.

[0071] Fitness calculation: After each update of the particle position, calculate the fitness value of the particle. The fitness value is the comprehensive result calculated by the objective function (energy consumption minimization) and the constraint condition (load demand satisfaction rate). The fitness function is as follows: f(p i ) = E total (p i ) + λ·max(0, 1 - load demand satisfaction rate); Among them, λ is a penalty coefficient used to penalize solutions where the load demand satisfaction rate does not reach 95%.

[0072] Iterative search: The particle swarm continuously iterates, searching for the optimal solution by updating the particle velocity and position until the preset stop condition is reached (such as the maximum number of iterations or the convergence of the objective function).

[0073] During the optimization process of the particle swarm optimization algorithm, the operating parameters of the air conditioning system will be dynamically adjusted, specifically including: Number of operating air conditioning units: Determine the number of air conditioning units to be started in each time period to ensure that the load demand is met while avoiding system overload.

[0074] Number of operating water pumps: Optimize the number of operating water pumps according to the cooling demand of the air conditioning system to reduce unnecessary energy consumption.

[0075] Set value of the outlet water temperature: Dynamically adjust the set value of the outlet water temperature according to the external meteorological conditions and the load demand of the air conditioning system to reduce energy consumption while meeting the load demand.

[0076] Operating frequency: Optimize the operating frequency of the air conditioning equipment according to the load prediction results and the real-time state of the air conditioning system to reduce energy consumption while ensuring the load demand.

[0077] The optimized operating parameters can be output in real time to guide the operation adjustment of the air conditioning system, so as to achieve the optimal energy efficiency in each time period.

[0078] After the optimization process is completed, the optimal parameters output by the particle swarm optimization algorithm will be applied to the air conditioning system in real time to generate an energy-saving operation plan that meets the load demand and has the optimal energy efficiency. This plan can dynamically adjust the operation strategy of the air conditioning system according to the changes in the actual load demand and meteorological conditions to ensure that the system is always in an energy-saving state.

[0079] In this embodiment, the operating parameters of the air conditioning system are dynamically optimized by the particle swarm optimization algorithm to ensure the minimization of the total energy consumption of the system while meeting the load demand. This method not only effectively reduces the energy consumption of the air conditioning system, but also can adapt to different load demands and external meteorological conditions, providing a real-time and energy-saving operation plan.

[0080] For step S5, in this embodiment, in order to ensure that the air-conditioning load prediction model continuously maintains high accuracy in a dynamic environment, a method of periodically or when the prediction error exceeds a set threshold to update the Gradient Boosting Decision Tree (GBDT) model and the Long Short-Term Memory Network (LSTM) model is proposed. This method combines a periodic update and an error-triggered update mechanism to cope with the impact of meteorological environment changes or seasonal fluctuations on the model performance, thereby improving the air-conditioning load prediction accuracy and energy-saving effect.

[0081] The update mechanism of the model mainly includes two triggering methods: Periodic update: As described above, every certain period (such as monthly or quarterly), the Gradient Boosting Decision Tree and the LSTM model are retrained. This periodic update ensures that the model can adapt to the impact of meteorological environment and seasonal fluctuations on load prediction.

[0082] Error threshold-triggered update: When the model prediction error exceeds the set threshold, an update is triggered. According to the above definition, when the relative error exceeds 10%, it indicates that the prediction accuracy of the current model has decreased, and the error needs to be corrected by retraining the model. The formula for calculating the error is as follows: where y pred is the predicted load value; y true is the actual load value; |·| represents the absolute value operation.

[0083] If the relative error exceeds 10%, the update of the Gradient Boosting Decision Tree and the LSTM model is triggered.

[0084] According to the above steps, the update process of the model is as follows: Data collection and preparation: When updating, the latest meteorological data, load prediction values, and other relevant features need to be collected, and data cleaning, feature selection, and preprocessing are performed. This step ensures that the input data used for model training has the latest and high-quality features.

[0085] GBDT model update: For the Gradient Boosting Decision Tree model, it is trained by an incremental learning method combined with new data to update the model structure and parameters. The update steps of the model include: Retrain the decision tree by minimizing the loss function (such as mean squared error) to gradually reduce the prediction error of the model; Update the structure and weights of the tree to ensure that each new tree can correct the error in the previous round of prediction; Optimize the hyperparameters (such as tree depth, learning rate, number of iterations, etc.) to ensure the best performance of the model under new data.

[0086] LSTM model update: For the LSTM model, the update process includes: By retraining the network, adjust the weights and biases of the network using new training data; Update the network structure, such as adjusting the number of layers of the LSTM or the number of neurons in each layer, to improve the adaptability of the model to new data; Use the standard backpropagation algorithm to optimize the network weights, and gradually optimize the model performance by calculating the loss function (such as mean squared error).

[0087] Evaluation and selection: After completing the updated training of the model, evaluate the performance of the updated model, and ensure its effectiveness on the new dataset through methods such as cross-validation. For evaluation metrics, continue to use standards such as mean squared error (MSE), and calculate the gap between the predicted value and the actual value. If the new model performs excellently, replace the original model.

[0088] Once the model update is completed, the new GBDT and LSTM models will be deployed to the real-time system to start air conditioning load prediction. By continuously updating the model, the system can maintain high-precision prediction ability in a changing environment, thereby providing a more accurate energy-saving scheduling scheme for the air conditioning system.

[0089] This embodiment ensures that the gradient boosting decision tree and long short-term memory network models can continuously adapt to meteorological environment changes and seasonal fluctuations through a periodic update and error threshold-triggered update mechanism. By retraining the model regularly or triggeringly, the accuracy of air conditioning load prediction is further improved, providing a more accurate prediction basis for the energy-saving scheduling of the air conditioning system.

[0090] The present invention also provides an air conditioning load prediction and energy-saving system based on machine learning, including: A multi-source data collection and preprocessing module for collecting historical building air conditioning load data, multi-dimensional meteorological data, and equipment operation parameters, and preprocessing the data; A feature selection module for inputting the historical load data and meteorological data into the gradient boosting decision tree model, calculating the importance scores of each meteorological feature, and screening key meteorological features whose cumulative score ratio reaches a preset threshold; A load prediction module for using the screened key meteorological features and historical load data as inputs, constructing a load prediction model using a long short-term memory network, and outputting the air conditioning load prediction value for a future period; An energy-saving optimization module for dynamically optimizing the operation parameters of the air conditioning system based on the particle swarm optimization algorithm; A model update module for periodically or when the prediction error exceeds a set threshold, updating the gradient boosting decision tree model and the long short-term memory network model The system of this embodiment can be used to execute the above method embodiment, and its principle and technical effects are similar, so they will not be elaborated here.

[0091] Although embodiments of the present invention have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An air-conditioning load prediction and energy-saving method based on machine learning, characterized in that, It includes the following steps: S1. Multi-source data collection, which includes collecting historical load data of building air conditioners, multi-dimensional meteorological data, and equipment operation parameters, and preprocessing the multi-source data; S2. Input the historical load data and meteorological data into the gradient boosting decision tree model, calculate the importance scores of each meteorological feature, and screen the key meteorological features whose cumulative score ratio reaches a preset threshold; S3. Use the screened key meteorological features and historical load data as inputs, construct a load prediction model using a long short-term memory network, and output the predicted value of the air conditioner load in the future period; S4. Dynamically optimize the operation parameters of the air conditioning system based on the particle swarm optimization algorithm; S5. Periodically or when the prediction error exceeds the set threshold, update the gradient boosting decision tree model and the long short-term memory network model.

2. The air-conditioning load prediction and energy-saving method based on machine learning according to claim 1, wherein In step S1, the multi-dimensional meteorological data includes temperature, humidity, wind speed, and sunshine intensity; the equipment operation parameters include the water inlet temperature of the unit, the water outlet temperature of the unit, the pump flow rate, and the number of operating units.

3. A machine learning-based air conditioning load prediction and energy saving method according to claim 1, characterized in that, In step S1, the data preprocessing includes: Clean the collected multi-source data to eliminate duplicate data and invalid data; Detect and delete outliers based on the isolation forest algorithm; Fill in the missing data using the time series interpolation method, where the time series interpolation method is the cubic spline interpolation method.

4. A machine learning-based air conditioning load prediction and energy saving method according to claim 1, characterized in that, In step S2, the preset threshold is 80%.

5. A machine learning-based air-conditioning load prediction and energy-saving method according to claim 1, characterized in that, In step S3, use the historical load data sampled in hours for the previous 24 hours and the screened key meteorological features as inputs, output the predicted value of the air conditioner load for the next 1 hour, and automatically adjust the hyperparameters of the long short-term memory network using an optimization algorithm.

6. A machine learning-based air conditioning load prediction and energy saving method according to claim 5, characterized in that, The hyperparameters include the number of hidden layers, the learning rate, and the Dropout rate, and the optimization algorithm is the whale optimization algorithm.

7. A method for air-conditioning load prediction and energy saving based on machine learning according to claim 1, characterized in that, Step S4 includes: Taking the load prediction value as the input, construct an objective function to minimize the total energy consumption of the air conditioning system; Under the constraint condition that the load demand satisfaction rate ≥ 95%, search for the Pareto optimal solution set through the particle swarm optimization algorithm; Dynamically optimize the operation parameters including the number of operating units of the unit, the number of operating units of the pump, the set value of the water outlet temperature, and the operating frequency; And real-time output an energy-saving operation plan that meets the optimization objectives.

8. A machine learning-based air conditioner load prediction and energy-saving method according to claim 1, characterized in that Step S5 includes: Periodically retrain the gradient boosting decision tree model and the long short-term memory network model to adapt to meteorological environment or seasonal changes; Trigger an update when the prediction error exceeds the set threshold, and the threshold is 10%.

9. An air-conditioning load prediction and energy-saving system based on machine learning, which is applied to an air-conditioning load prediction and energy-saving method based on machine learning according to any one of claims 1-8, and is characterized in that, It includes: A multi-source data collection and preprocessing module, which is used to collect historical load data of building air conditioners, multi-dimensional meteorological data, and equipment operation parameters, and preprocess the data; A feature selection module, which is used to input the historical load data and meteorological data into the gradient boosting decision tree model, calculate the importance scores of each meteorological feature, and screen the key meteorological features whose cumulative score ratio reaches a preset threshold; A load prediction module, which is used to use the screened key meteorological features and historical load data as inputs, construct a load prediction model using a long short-term memory network, and output the predicted value of the air conditioner load in the future period; An energy-saving optimization module, which is used to dynamically optimize the operating parameters of the air-conditioning system based on the particle swarm optimization algorithm; A model update module, which is used to update the gradient boosting decision tree model and the long short-term memory network model periodically or when the prediction error exceeds a set threshold.

Images