A water affair private cloud load prediction method
By employing the CNN-BiLSTM-Attention model and the INRBO algorithm to optimize hyperparameters in a private water utilities cloud, the accuracy and stability issues of the water utilities load prediction model were resolved, enabling accurate prediction and efficient utilization of water utilities cloud resources.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHONGQING YUANTONG ELECTRONICS TECH DEV CO LTD
- Filing Date
- 2026-01-06
- Publication Date
- 2026-05-29
Smart Images
Figure CN122120166A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of cloud resource management technology, specifically to a method for load prediction in a private water cloud. Background Technology
[0002] The water industry is an industrial chain comprised of raw water, water supply, water conservation, drainage, sewage treatment, and water resource recycling. With the acceleration of digital transformation and the continuous advancement of smart water management, the digital and intelligent transformation of the water industry is deepening. The cloud-based deployment of various water business systems (such as Supervisory Control and Data Acquisition (SCADA) systems, big data analysis systems, DMA (District Metering Area) metering systems, hydraulic model simulation systems, online customer service response systems, production operation control platforms, integrated dispatching systems, and marketing management systems) has become an inevitable trend for improving management efficiency and optimizing service quality. However, due to the special nature of water data and the stringent security requirements, water companies prioritize building dedicated private cloud platforms rather than using public cloud services. The reason is that water data covers key information such as raw water quality, water supply network layout, user water usage information, and sewage treatment process parameters. Some data is related to urban infrastructure security and public privacy, and is considered sensitive core data. Private clouds can achieve autonomous control of core data, avoid the risks of data leakage and unauthorized access in the public cloud environment through personalized security protection strategies, and at the same time ensure the continuity of business services, avoiding service interruptions caused by external network environment, service provider operation and maintenance strategies, and other factors.
[0003] In the actual operation of a private cloud platform, resource utilization must closely match the dynamic needs of water-related businesses. Water-related businesses exhibit significant periodicity and volatility. Scenarios such as peak water usage periods, extreme weather, handling of sudden water-related incidents, and batch task execution in business systems can lead to a surge in business load, while off-peak periods pose a risk of resource idleness. This places flexible and adaptable requirements on the configuration of computing, storage, and network resources in the private cloud. To achieve efficient utilization of private cloud resources, reduce energy consumption, and ensure the stable operation of various water-related businesses, accurate cloud load forecasting based on business needs is necessary. By analyzing and modeling historical load data, the load change trends at different times in the future can be predicted, providing a basis for decision-making on the dynamic scheduling of private cloud resources. This allows for advance resource expansion or contraction adjustments, avoiding problems such as resource allocation imbalances, service lag, imbalances between service quality assurance and economic efficiency, lack of forward-looking input for intelligent operation and maintenance decisions, and difficulty in proactively mitigating system risks caused by load fluctuations. Therefore, in the context of building a smart water private cloud, conducting load forecasting research based on business needs is a key support for optimizing the use of private cloud resources and ensuring efficient business operation.
[0004] Load forecasting, as a forward-looking core technology in cloud resource management, plays a crucial role in achieving efficient cloud resource utilization, reducing energy consumption, improving user satisfaction, and ensuring service quality. Its core logic lies in achieving accurate perception of cloud service demand through in-depth analysis and modeling of historical load data, thereby predicting resource usage trends in future periods. This provides scientific support for scheduling decisions, ultimately enabling advance adjustments to resource allocation, effectively avoiding host overload, reducing unnecessary virtual machine migrations, and improving the operational stability and energy efficiency of the water cloud platform. Traditional cloud load forecasting models are based on statistical principles and machine learning. Many cloud forecasting models are based on statistical principles, such as MA and ARMA. However, their drawback is insufficient extraction of complex load features, resulting in poor prediction accuracy. Machine learning-based cloud forecasting models include SVM and BRR. While they offer improved prediction capabilities compared to traditional statistical methods, they still have shortcomings in extracting features from nonlinear water load data, and their prediction accuracy and ability to model long-term time-series dependencies still need improvement. Summary of the Invention
[0005] The present invention aims to provide a water private cloud load prediction method, which can achieve relatively accurate water load prediction in complex and dynamic cloud environments.
[0006] To address the aforementioned technical problems, the present invention provides the following technical solution: This invention provides a method for load forecasting in a private water utilities cloud, comprising the following steps: Preprocess water load data; The CNN-BiLSTM-Attention model was used to complete data feature extraction and temporal modeling. The INRBO algorithm was used to optimize the model hyperparameters. The model performs load forecasting.
[0007] By adopting the above technical solution, a data processing and modeling framework for cloud physical host load prediction is constructed. A deep learning model combining a convolutional neural network (CNN) with a SE attention mechanism and a bidirectional long short-term memory network (BiLSTM) is employed to fully extract the data features of cloud load. Simultaneously, an improved intelligent optimization algorithm, INRBO, is used to globally search and optimize the model's hyperparameters, thereby obtaining the optimal combination of training parameters and improving the model's prediction performance. This invention fully considers the complexity and volatility of load sequences in the cloud environment and proposes a cloud load prediction model, INRBO-CNN-BiLSTM-Attention, which integrates feature enhancement and parameter optimization mechanisms. This model can effectively capture load change trends, significantly improve prediction accuracy and stability, and is well-suited to real-world cloud computing scenarios.
[0008] Optionally, the preprocessed water utilities private cloud load data includes: First, feature engineering enhancement is performed, followed by data cleaning operations such as missing value imputation, outlier removal, and smoothing filtering. Then, the processed dataset is divided into training, validation, and test sets. Finally, advanced feature engineering enhancement is performed on the training set.
[0009] Optionally, the CNN-BiLSTM-Attention model includes: Convolutional layers are used to extract local features; SE attention mechanism is used to enhance the representation of important features; BiLSTM time series modeling layer is used to capture bidirectional dependencies in time series; A fully connected output layer for final load prediction output.
[0010] Optionally, the optimization of model hyperparameters using the INRBO algorithm includes: Initialize the population size N, the maximum number of iterations MaxIt, the upper and lower bounds of the variables LB and UB, the problem dimension dim, and the determinant DF. The objective function is defined as the mean squared error of the CNN-BiLSTM-Attention model on the validation set. Generate an initial population, where each individual represents a set of hyperparameter combinations, including the learning rate lr, the number of hidden layer nodes hd, and the L2 regularization coefficient l2; Perform multiple rounds of iterative optimization operations.
[0011] Optionally, the population is initialized using a refraction-based inverse learning strategy with the INRBO algorithm-optimized model, and the specific steps are as follows: Generate a random initial solution set: ; Calculate the corresponding reverse refraction solution: ; Merge X1 and X2 to construct the initial population, and prune or supplement when the population size is mismatched to ensure that the total number of individuals is N.
[0012] Optionally, the multi-round iterative optimization operation includes: Update the decay factor; Preserve the elite individuals in the current population; For each individual i, the solution is updated, including: Randomly select two different individuals to construct a direction vector; Candidate solutions are generated based on NRSR search rules; The candidate solution is mixed with the mean of individuals in the elite pool using cross-probability to generate an updated solution; If the dynamic factor judgment condition is met, a perturbation solution is generated and the current solution is updated; The newly generated solutions are restricted to the upper and lower bounds of the variables and passed to the CNN-BiLSTM-Attention model as training hyperparameters.
[0013] Optionally, generating candidate solutions based on NRSR search rules includes: calculate and The value is shown in the following formula: In the formula, This represents a normally distributed random number with a mean of 0 and a variance of 1. Indicates the optimal position; Indicates the worst position; In the formula, and yes Random numbers between; and These are different integers randomly selected from the population; and This represents two random individuals within a population; Update position variable , , As shown in the following formula: In the formula, The best position; Obtain candidate solutions As shown in the following formula: .
[0014] Optionally, a simulated annealing acceptance criterion is introduced during the perturbation solution generation stage, as follows: If the fitness of the new solution is less than or equal to that of the current best solution, or if the probability is satisfied... If the new solution is accepted, the corresponding individuals in the elite pool will be updated.
[0015] Optionally, the following steps are performed after each iteration: The CNN-BiLSTM-Attention model is configured with hyperparameters corresponding to the current individual, trained on the training set, and used to make predictions on the validation set. The mean squared error (MSE) between the predicted and actual values on the validation set is calculated and used as the fitness value of the individual. If the new solution is better than the current individual's fitness, then update the individual's position and fitness; If the new solution is better than the global optimal solution, then update the global optimal solution and its optimal position, and update the elite database simultaneously. If the new solution is worse than the current worst solution, then update the worst solution and its corresponding position.
[0016] Optionally, after completing one round of iteration, the population is sorted by fitness and the global optimal fitness value is updated; Once the maximum number of iterations MaxIt is reached, the optimal hyperparameter combination [lrbest, hdbest, l2best] is output.
[0017] In summary, the present invention has at least the following beneficial technical effects: This invention takes into account the complexity and volatility of water utilities' private cloud load and the need to predict various cloud resource indicators. It proposes an INRBO-based cloud load prediction method, which can achieve accurate prediction of cloud load and performs well on various cloud resource indicators, demonstrating good generalization performance. Attached Figure Description
[0018] Figure 1 This is a schematic diagram of the cloud load forecasting method.
[0019] Figure 2 This is a network structure diagram of the cloud load prediction model. Detailed Implementation
[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.
[0021] This invention provides a method for predicting the load of a private cloud for water services.
[0022] A method for load forecasting in a private water utilities cloud includes the following steps: S1: Preprocess water utility private cloud load data.
[0023] S2: Use the CNN-BiLSTM-Attention model to complete data feature extraction and temporal modeling.
[0024] The CNN-BiLSTM-Attention model is a hybrid deep learning model that integrates convolutional neural networks (CNN), bidirectional long short-term memory networks (BiLSTM), and attention mechanisms.
[0025] A two-layer one-dimensional convolutional neural network is used to extract local features from cloud water load data. The basic feature extraction layer uses a 3×1 convolutional kernel with 32 channels, while the deep feature extraction layer also uses a 3×1 convolutional kernel with 64 channels to capture higher-dimensional feature representations. After each convolutional layer, a ReLU activation function and a batch normalization layer are introduced to enhance the model's nonlinear expression ability, accelerate convergence, and reduce the risk of overfitting.
[0026] After the feature extraction layer subnetwork, an SE attention mechanism is added to dynamically weight channel features with a 1 / 4 compression ratio. First, global average pooling is performed on the feature map to generate a fixed-length feature vector, obtaining the global receptive field. Then, two fully connected layers are used to generate dynamic weight coefficients. Next, Sigmoid activation is used to achieve adaptive weighting of feature channels. Finally, the generated channel weights are multiplied with the original feature map channel by channel to enhance important features and suppress unimportant features, thus completing the channel attention-based recalibration of the original features.
[0027] Finally, a bidirectional long short-term memory network is used to perform time-series modeling on the extracted feature sequences. By jointly learning forward and backward time-series information, the model can comprehensively capture the long-term dependencies and change patterns in cloud water load data, thereby improving the accuracy and stability of the model's prediction of future loads.
[0028] S3: Optimize model hyperparameters using the INRBO algorithm.
[0029] The INRBO algorithm is an improved Newton-Raphson optimizer algorithm (NRBO).
[0030] Initialize the dynamic factor DF, refractive index parameter k, and the number of elite individuals EliteNum.
[0031] Input population size N, maximum number of iterations MaxIt, upper and lower bounds of variables LB and UB, determinant DF, dimension dim, and objective function fobj.
[0032] The population is initialized using a refraction-based inverse learning strategy, and the specific steps are as follows: (1) Generate a random initial solution set: ; (2) Calculate the corresponding reverse refraction solution: ; (3) Merge X1 and X2 to construct the initial population, and prune or supplement when the population size is mismatched to ensure that the total number of individuals is N; Calculate the fitness value of each individual in the initial population, sort by fitness, select the current best and worst solutions, and construct the initial elite pool.
[0033] Enter the iterative optimization stage. From iteration number 1 to MaxIt, perform the following operations: Furthermore, at the beginning of each iteration, according to the formula Update the attenuation factor; Reserve the top EliteNum optimal individuals in the current population as the elite library.
[0034] Perform the following operations on each individual i: Randomly select two different individuals to construct a direction vector; generate a candidate solution based on the NRSR search rule; use the crossover probability to mix the candidate solution with the mean of the individuals in the elite library to generate an updated solution; if the dynamic factor judgment condition rand < DF is satisfied, generate a perturbation solution according to the preset formula and update the current solution, otherwise keep the original solution unchanged.
[0035] Limit the newly generated solution within the variable upper and lower bounds and calculate its fitness; if the new solution is better than the current individual fitness, update the position and fitness of this individual; if the new solution is better than the global optimal solution, update the global optimal solution and the optimal position, and synchronously update the elite library; if the new solution is worse than the current worst solution, update the worst solution and its corresponding position.
[0036] Meanwhile, introduce the simulated annealing acceptance criterion: if the fitness of the new solution is less than or equal to the current optimal solution, or satisfies the probability , then accept the new solution and update the corresponding individual in the elite library; after completing one round of iteration, sort the population according to fitness and update the global optimal fitness value; when the maximum number of iterations is reached, output the optimal solution and its corresponding parameter combination.
[0037] The method for training the model is as follows: The INRBO part outputs the optimized hyperparameter combination, including the number of hidden layer nodes, the initial learning rate, and the L2 regularization coefficient; The BiLSTM-Attention part uses these parameters for network training and calculates the mean square error between the predicted value and the actual value on the validation set as the fitness value; Feed back the fitness value to the INRBO part to guide the population update, gradually approximate the optimal parameter combination through iterative optimization, realize the parameter optimization of the model, and improve the parameter tuning efficiency and prediction accuracy.
[0038] S4: The model performs load prediction.
[0039] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.
[0040] A method for predicting the load of a water service private cloud, as Figure 1 shown, the specific steps are as follows: 1) Obtain historical resource usage data of physical hosts in the cloud environment. This historical resource usage data is also known as water load data. It is in time series form and includes indicators such as physical machine CPU load, memory load, virtual machine CPU load, and memory load.
[0041] 2) Preprocess the water load data, including missing value imputation, outlier removal and smoothing filtering; then set the historical time step kim and the prediction step zim.
[0042] 3) Divide the dataset into training, validation, and test sets, and perform advanced feature engineering on the training set.
[0043] 4) Construct the CNN-BiLSTM-Attention network structure, such as... Figure 2 As shown, it includes convolutional layers, SE attention mechanism, BiLSTM temporal modeling layer and fully connected output layer.
[0044] 5) Set the initial parameters of the INRBO optimization algorithm, including population size N, maximum number of iterations MaxIt, upper and lower bounds of variables LB and UB, problem dimension dim, determination factor DF, and objective function fobj, where the objective function is defined as the mean squared error of the CNN-BiLSTM-Attention model on the validation set.
[0045] 6) Run the INRBO algorithm to generate the initial population. Each individual represents a set of hyperparameter combinations, including the learning rate lr, the number of hidden layer nodes hd, and the L2 regularization coefficient l2.
[0046] 7) From the it=1th round to MaxIt, perform iterative optimization operations.
[0047] Furthermore, the specific steps of step 7) are as follows: 7-1) Update the decay factor according to the current iteration round. . 7-2) Retain the elite individuals in the current population, with a number of EliteNum. 7-3) Update the solution for each individual i.
[0048] Furthermore, the specific steps of step 7-3) are as follows: 7-3-1) Randomly select two different individuals to construct direction vectors, and the current position vector is represented as... . 7-3-2) Generating candidate solutions based on NRSR search rules .
[0049] Furthermore, the specific steps of step 7-3-2) are as follows: 7-3-2-1) Calculation and values, as shown in the following formula: In the formula, represents a random number of a normal distribution with a mean of 0 and a variance of 1; represents the best position; represents the worst position.
[0050] In the formula, and are random numbers between; and are different integers randomly selected from the population; and represent two random individuals among the population.
[0051] 7-3-2-2) Update the position variables , , , as shown in the following formula: In the formula, is the best position.
[0052] 7-3-2-3) Obtain the candidate solution , as shown in the following formula: 7-3-3) Use the crossover probability to mix the candidate solution with the mean of the individuals in the elite library to generate an updated solution. The crossover rate decay formula is as follows: 7-3-4) If the dynamic factor judgment condition rand < DF is satisfied, generate a perturbation solution and update the current solution according to the following formula: In the formula: represents the best position, is a random number between, and respectively represent and random numbers between; is the mean function; This represents the position vector of all individuals in the current iteration. and These are random numbers, expressed by the following formulas: If rand > DF, then keep the original solution unchanged. 7-3-5) Restrict the newly generated solution to the upper and lower bounds of the variables and pass it to the CNN-BiLSTM-Attention model as training hyperparameters.
[0053] 7-4) Use the hyperparameter configuration corresponding to the current individual to configure the CNN-BiLSTM-Attention model, train it on the training set, and make predictions on the validation set. 7-5) Calculate the mean squared error (MSE) between the predicted and true values on the validation set, and use this as the fitness value for the individual. 7-6) If the new solution is better than the current individual's fitness, update the individual's position and fitness. 7-7) If the new solution is better than the global optimum, update the global optimum and its optimal position, and simultaneously update the elite pool. 7-8) If the new solution is worse than the current worst solution, update the worst solution and its corresponding position. 7-9) Introduce the simulated annealing acceptance criterion: if the fitness of the new solution is less than or equal to the current best solution, or if the probability is satisfied... If the new solution is accepted, the corresponding individuals in the elite pool will be updated.
[0054] 8) After completing one round of iteration, sort the population by fitness and update the global optimal fitness value.
[0055] 9) When the maximum number of iterations is reached, output the optimal hyperparameter combination [lrbest, hdbest, l2best].
[0056] 10) Retrain the model using the optimal hyperparameters and make the final prediction on the test set, and calculate the evaluation metric of the prediction results.
[0057] Experiment 1: The Alibaba Cluster-Trace-v2018 dataset was used for testing. CPU water load data from physical machines was selected as the prediction target, and the proposed method was applied for prediction. The method was compared with the basic CNN-BiLSTM model. To accurately evaluate the performance of the prediction algorithm, this experiment used Mean Absolute Error (MAE), Root Mean Square Error (RMSE), Mean Absolute Percentage Error (MAPE), and the coefficient of determination. These are the four evaluation metrics: (R Squared).
[0058] Table 1 compares this invention with the CNN-BiLSTM model. The results are shown in Table 1. The method of this invention demonstrates superior prediction performance across all metrics. Compared to the CNN-BiLSTM model, the method of this invention shows reductions in MAE, RMSE, and MAPE, indicating smaller prediction errors. In terms of R², the method of this invention improves by 4.2% compared to CNN-BiLSTM, closer to the ideal value of 1, indicating a better fit to the training data. Considering the overall performance across all metrics, in predicting physical host CPU water load data, the method of this invention outperforms the CNN-BiLSTM model in prediction accuracy, stability, and explanatory power, demonstrating more reliable prediction capabilities.
[0059] Experiment 2: Using the Alibaba Cluster-Trace-v2018 dataset, the memory load of the physical machine was selected as the prediction target, and the method of this invention was applied for prediction. The proposed method was compared with the basic model CNN-BiLSTM.
[0060] Table 2 compares this invention with the CNN-BiLSTM model. MAE RMSE MAPE <![CDATA[R 2 ]]> Method of the present invention 0.45480 0.70479 0.55857 0.92331 CNN-BiLSTM 1.12441 1.55318 1.33073 0.80106 The results are shown in Table 2. Similarly, the MAE, RMSE and MAPE of the method of this invention are reduced by more than 50% compared with CNN-BiLSTM, while R² is significantly improved from 0.80106 to 0.92331, demonstrating the excellent predictive performance of the improved model.
[0061] The above description of the embodiments is only used to provide a detailed introduction to the technical solution of the present invention. However, the description of the above embodiments is only for the purpose of helping to understand the method and core idea of the present invention, and should not be construed as a limitation of the present invention. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be covered within the protection scope of the present invention.
Claims
1. A method for load forecasting in a water utility private cloud, characterized in that, Includes the following steps: Preprocess water utilities' private cloud load data; The CNN-BiLSTM-Attention model was used to complete data feature extraction and temporal modeling. The INRBO algorithm was used to optimize the model hyperparameters. The model performs load forecasting.
2. The water utility private cloud load forecasting method as described in claim 1, characterized in that, The pre-processed water utilities private cloud load data includes: First, feature engineering enhancement is performed, followed by data cleaning operations such as missing value imputation, outlier removal, and smoothing filtering. Then, the processed dataset is divided into training, validation, and test sets. Finally, advanced feature engineering enhancement is performed on the training set.
3. The water utility private cloud load forecasting method as described in claim 1, characterized in that, The CNN-BiLSTM-Attention model includes: Convolutional layers are used to extract local features; SE attention mechanism is used to enhance the representation of important features; BiLSTM time series modeling layer is used to capture bidirectional dependencies in time series; A fully connected output layer for final load prediction output.
4. The water utility private cloud load forecasting method as described in claim 1, characterized in that, The optimization of model hyperparameters using the INRBO algorithm includes: Initialize the population size N, the maximum number of iterations MaxIt, the upper and lower bounds of the variables LB and UB, the problem dimension dim, and the determinant DF. The objective function is defined as the mean squared error of the CNN-BiLSTM-Attention model on the validation set. Generate an initial population, where each individual represents a set of hyperparameter combinations, including the learning rate lr, the number of hidden layer nodes hd, and the L2 regularization coefficient l2; Perform multiple rounds of iterative optimization operations.
5. The water utility private cloud load forecasting method as described in claim 4, characterized in that, The specific steps for initializing the population using the INRBO algorithm-optimized model and the refraction-backward learning strategy are as follows: Generate a random initial solution set: ; Calculate the corresponding reverse refraction solution: ; Merge X1 and X2 to construct the initial population, and prune or supplement when the population size is mismatched to ensure that the total number of individuals is N.
6. The water utilities private cloud load forecasting method as described in claim 4, characterized in that, The multi-round iterative optimization operation includes: Update the decay factor; Preserve the elite individuals in the current population; For each individual i, the solution is updated, including: Randomly select two different individuals to construct a direction vector; Candidate solutions are generated based on NRSR search rules; The candidate solution is mixed with the mean of individuals in the elite pool using cross-probability to generate an updated solution; If the dynamic factor judgment condition is met, a perturbation solution is generated and the current solution is updated; The newly generated solutions are restricted to the upper and lower bounds of the variables and passed to the CNN-BiLSTM-Attention model as training hyperparameters.
7. The water utility private cloud load forecasting method as described in claim 6, characterized in that, The generation of candidate solutions based on NRSR search rules includes: calculate and The value is shown in the following formula: ， In the formula, This represents a normally distributed random number with a mean of 0 and a variance of 1. Indicates the optimal position; Indicates the worst position; ， In the formula, and yes Random numbers between; and These are different integers randomly selected from the population; and This represents two random individuals within a population; Update position variable , , As shown in the following formula: ， ， ， In the formula, The best position; Obtain candidate solutions As shown in the following formula: 。 8. The water utility private cloud load forecasting method as described in claim 6, characterized in that, A simulated annealing acceptance criterion is introduced during the perturbation solution generation stage, as follows: If the fitness of the new solution is less than or equal to that of the current best solution, or if the probability is satisfied... If the new solution is accepted, the corresponding individuals in the elite pool will be updated.
9. The water utility private cloud load forecasting method as described in claim 4, characterized in that, After each iteration, perform the following steps: The CNN-BiLSTM-Attention model is configured with hyperparameters corresponding to the current individual, trained on the training set, and used to make predictions on the validation set. The mean squared error (MSE) between the predicted and actual values on the validation set is calculated and used as the fitness value of the individual. If the new solution is better than the current individual's fitness, then update the individual's position and fitness; If the new solution is better than the global optimal solution, then update the global optimal solution and its optimal position, and update the elite database simultaneously. If the new solution is worse than the current worst solution, then update the worst solution and its corresponding position.
10. The water utilities private cloud load forecasting method as described in claim 4, characterized in that, After completing one round of iterations, the population is sorted by fitness and the global optimal fitness value is updated. Once the maximum number of iterations MaxIt is reached, the optimal hyperparameter combination [lrbest, hdbest, l2best] is output.