Computing power demand prediction system using deep learning

Abstract

The invention provides a computing power demand prediction system using deep learning, and relates to the field of electrical digital data processing, and the system comprises a data collection module, a feature extraction module, a model training module and a prediction execution module, the feature extraction module is used for extracting feature information from sample data, the model training module is used for carrying out deep learning training on the sample data, and the prediction execution module is used for predicting the computing power demand of a task based on a learning result; according to the system, the prediction accuracy of the model is improved through deep learning training, the computing power demand of a task is output, and data support is provided for subsequent resource allocation.

Description

Technical Field

[0001] The present invention relates to the field of electrical digital data processing, and in particular to a computing power demand prediction system using deep learning. Background Art

[0002] With the rapid development of cloud computing, big data, and artificial intelligence technologies, the demand for computing resources is growing exponentially. Especially in deep learning and data-intensive applications, it is crucial to rationally configure and efficiently manage computing resources. The complexity and scale of current computing tasks are increasing, and the demand for computing resources fluctuates greatly and is unpredictable. If the computing power allocation cannot be dynamically adjusted according to the actual needs of the task, it is easy to cause resource waste or performance bottlenecks. Therefore, establishing a system that can accurately predict computing power requirements can significantly improve resource utilization and task response speed, and provide data support for resource planning and load balancing.

[0003] The foregoing discussion of the background art is intended only to facilitate an understanding of the present invention. This discussion does not acknowledge or admit that any of the material referred to is part of the common general knowledge.

[0004] Now many computing power prediction systems have been developed. After a lot of searching and reference, it is found that the existing computing power prediction systems are such as the system disclosed in the publication number CN117952669B. These system methods generally include: counting the amount of data generated by individuals in the target area, the amount of data generated by pan-entertainment, and the amount of data generated by sensor devices; based on the amount of data generated by individuals, the amount of data generated by pan-entertainment, and the amount of data generated by sensor devices, using the computing power demand equivalent conversion method to obtain the computing power demand of the target area; calculating the historical computing power demand sequence; using the twin neural network to conduct deep learning on the historical computing power demand sequence to obtain the computing power demand prediction network. However, during the model training process, this system can only process one sample at a time, and the training efficiency needs to be improved. Summary of the invention

[0005] The purpose of the present invention is to propose a computing power demand prediction system using deep learning to address the existing deficiencies.

[0006] The present invention adopts the following technical solution: A computing power demand prediction system using deep learning, including a data collection module, a feature extraction module, a model training module and a prediction execution module; The data collection module is used to collect sample data for training, the feature extraction module is used to extract feature information from the sample data, the model training module is used to perform deep learning training on the sample data, and the prediction execution module predicts the computing power requirements of the task based on the learning results; The data collection module includes a task collection unit, a data preprocessing unit and a sample storage unit. The task collection unit is used to collect the task data after execution. The data preprocessing unit is used to uniformly process the task data into data in the form of samples. The sample storage unit is used to encode and store the samples. The feature extraction unit includes a feature selection unit, a feature engineering unit and a feature conversion unit, wherein the feature selection unit is used to select key features related to computing power requirements, the feature engineering unit is used to construct high-order features from the selected features, and the feature conversion unit is used to convert the constructed features into a format supported by the model; The model training module includes a model construction unit, a learning and training unit and a model optimization unit. The model construction unit is used to construct a neural network structure to generate a training model. The learning and training unit is used to control the training model to perform learning and training on sample data. The model optimization unit optimizes the training model based on the training results. The prediction execution module includes a model storage unit, a real-time prediction unit and a feedback adjustment unit. The model storage unit is used to store model data for prediction. The real-time prediction unit is used to receive prediction tasks and use the model to make predictions. The feedback adjustment unit is used to collect actual operation data of the task, process the error information obtained, and then feed it back to the data collection module. Further, the feature selection unit includes an initial screening processor, a statistical analysis processor and a feature evaluation processor, wherein the initial screening processor is used to perform a simple screening and filtering of features that are not universal and obviously irrelevant, the statistical analysis processor is used to analyze the correlation between each feature and the computing power requirement, and the feature evaluation processor is used to evaluate the importance of the feature to the computing power requirement; The feature evaluation processor sorts the features from large to small according to the associated values, and calculates the importance index E of the features according to the following formula: ; Among them, u is the sorting number of the feature, n is the number of features, and r is the associated value; Furthermore, the feature engineering unit includes a feature interaction generation processor, a multidimensional feature aggregation processor and an engineering record processor, wherein the feature interaction generation processor is used to interactively process multiple features to generate new deep features, the multidimensional feature aggregation processor is used to select multiple deep features to form a feature vector, and the engineering record processor is used to record the interaction method and the vector formation method; The feature interaction generation processor interactively processes the features according to the following formula to obtain the deep feature X: ; Among them, H i() represents the i-th interaction function, x(1) and x(2) represent two interaction features, E(1) represents the importance index of the first interaction feature, and E(2) represents the importance index of the second interaction feature; Further, the learning and training unit includes a sample input processor, a training feedback processor and a training control processor, wherein the sample input processor is used to obtain sample data and place the corresponding feature vector at the correct position of the input layer, the training feedback processor is used to perform feedback statistics on the calculation status of each node in the hidden layer, and the training control processor controls the output state of the nodes in the hidden layer based on the statistical data; Furthermore, the training control processor calculates the congestion index P of each node according to the following formula: ; Among them, a is the range length, N0 is the count value of this node, N i is the count value of the i-th subsequent node; When the congestion index is less than the threshold, the node can directly send the calculation result to the next node, otherwise the output channel is closed until the congestion index is less than the threshold.

[0007] The beneficial effects achieved by the present invention are: This system can effectively improve the learning quality of the model by screening and reconstructing features, and output computing power requirements that are closer to the actual situation. During the training process, the transmission output of the node is controlled, so that multiple samples can be trained and processed in a single model at the same time without disturbing the data. Compared with the traditional method of processing one sample before processing the next sample, it can effectively improve the training efficiency of the model.

[0008] To further understand the features and technical contents of the present invention, please refer to the following detailed description and drawings of the present invention. However, the drawings provided are only for reference and description and are not intended to limit the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] Figure 1 It is a schematic diagram of the overall structural framework of the present invention; Figure 2 This is a schematic diagram of the data collection module of the present invention; Figure 3 It is a schematic diagram of the feature extraction unit of the present invention; Figure 4 This is a schematic diagram of the model training module of the present invention; Figure 5 This is a schematic diagram of the structure of the predictive execution module of the present invention; Figure 6 It is a comparison diagram of the training effect of the present invention. DETAILED DESCRIPTION

[0010] The following is an explanation of the embodiments of the present invention through specific embodiments. Those skilled in the art can understand the advantages and effects of the present invention from the contents disclosed in this specification. The present invention can be implemented or applied through other different specific embodiments, and the details in this specification can also be modified and changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. In addition, the drawings of the present invention are only simple schematic illustrations and are not depicted according to actual sizes. It is stated in advance. The following embodiments will further explain the relevant technical contents of the present invention in detail, but the disclosed contents are not intended to limit the scope of protection of the present invention.

[0011] Embodiment 1: This embodiment provides a computing power demand prediction system using deep learning, combined with Figure 1 , including data collection module, feature extraction module, model training module and prediction execution module; The data collection module is used to collect sample data for training, the feature extraction module is used to extract feature information from the sample data, the model training module is used to perform deep learning training on the sample data, and the prediction execution module predicts the computing power requirements of the task based on the learning results; The data collection module includes a task collection unit, a data preprocessing unit and a sample storage unit. The task collection unit is used to collect the task data after execution. The data preprocessing unit is used to uniformly process the task data into data in the form of samples. The sample storage unit is used to encode and store the samples. The feature extraction unit includes a feature selection unit, a feature engineering unit and a feature conversion unit, wherein the feature selection unit is used to select key features related to computing power requirements, the feature engineering unit is used to construct high-order features from the selected features, and the feature conversion unit is used to convert the constructed features into a format supported by the model; The model training module includes a model construction unit, a learning and training unit and a model optimization unit. The model construction unit is used to construct a neural network structure to generate a training model. The learning and training unit is used to control the training model to perform learning and training on sample data. The model optimization unit optimizes the training model based on the training results. The prediction execution module includes a model storage unit, a real-time prediction unit and a feedback adjustment unit. The model storage unit is used to store model data used for prediction. The real-time prediction unit is used to receive prediction tasks and use the model to make predictions. The feedback adjustment unit is used to collect actual running data of the task, process the error information obtained, and then feed it back to the data collection module.

[0012] The feature selection unit includes an initial screening processor, a statistical analysis processor and a feature evaluation processor, wherein the initial screening processor is used to perform simple screening and filtering of features that are not universal and obviously irrelevant, the statistical analysis processor is used to analyze the correlation between each feature and the computing power requirement, and the feature evaluation processor is used to evaluate the importance of the feature to the computing power requirement; The feature evaluation processor sorts the features from large to small according to the associated values, and calculates the importance index E of the features according to the following formula: ; Among them, u is the sorting number of the feature, n is the number of features, and r is the associated value.

[0013] The feature engineering unit includes a feature interaction generation processor, a multidimensional feature aggregation processor and an engineering record processor, wherein the feature interaction generation processor is used to interactively process multiple features to generate new deep features, the multidimensional feature aggregation processor is used to select multiple deep features to form a feature vector, and the engineering record processor is used to record the interaction method and the vector formation method; The feature interaction generation processor interactively processes the features according to the following formula to obtain the deep feature X: ; Among them, H i () represents the i-th interaction function, x(1) and x(2) represent two interaction features, E(1) represents the importance index of the first interaction feature, and E(2) represents the importance index of the second interaction feature.

[0014] The learning and training unit includes a sample input processor, a training feedback processor and a training control processor. The sample input processor is used to obtain sample data and place the corresponding feature vector at the correct position of the input layer. The training feedback processor is used to perform feedback statistics on the calculation status of each node in the hidden layer. The training control processor controls the output state of the nodes in the hidden layer based on the statistical data.

[0015] The training control processor calculates the crowding index P of each node according to the following formula: ; Among them, a is the range length, N0 is the count value of this node, N i is the count value of the i-th subsequent node; When the congestion index is less than the threshold, the node can directly send the calculation result to the next node, otherwise the output channel is closed until the congestion index is less than the threshold. Example

[0016] This embodiment includes all the contents of the first embodiment, and provides a computing power demand prediction system using deep learning, including a data collection module, a feature extraction module, a model training module and a prediction execution module.

[0017] The data collection module is used to collect sample data for training, the feature extraction module is used to extract feature information from the sample data, the model training module is used to perform deep learning training on the sample data, and the prediction execution module predicts the computing power requirements of the task based on the learning results.

[0018] Combination Figure 2 The data collection module includes a task collection unit, a data preprocessing unit and a sample storage unit. The task collection unit is used to collect task data that has been executed, the data preprocessing unit is used to uniformly process the task data into data in the form of samples, and the sample storage unit is used to encode and store the samples.

[0019] Combination Figure 3 The feature extraction unit includes a feature selection unit, a feature engineering unit and a feature conversion unit. The feature selection unit is used to select key features related to computing power requirements, the feature engineering unit is used to construct high-order features from the selected features, and the feature conversion unit is used to convert the constructed features into a format supported by the model.

[0020] Combination Figure 4 The model training module includes a model construction unit, a learning and training unit and a model optimization unit. The model construction unit is used to construct a neural network structure to generate a training model. The learning and training unit is used to control the training model to learn and train sample data. The model optimization unit optimizes the training model based on the training results.

[0021] Combination Figure 5 The prediction execution module includes a model storage unit, a real-time prediction unit and a feedback adjustment unit. The model storage unit is used to store model data for prediction. The real-time prediction unit is used to receive prediction tasks and use the model to make predictions. The feedback adjustment unit is used to collect actual running data of the task, process the error information obtained, and then feed it back to the data collection module.

[0022] The task acquisition unit includes an acquisition source manager, a task receiving processor and a task verification processor. The acquisition source manager is used to manage the address information of the task source, the task receiving processor is used to receive task data from the task source address, and the task verification processor is used to verify the necessary information in the task data and filter the task data.

[0023] The data preprocessing unit includes a format information register, a format conversion processor and a sample output processor. The format information register is used to store the format information of the sample. The format conversion processor is used to convert the task data into sample format information. The sample output processor is used to output the converted sample to the sample storage unit.

[0024] The sample storage unit includes a sample encoding processor, a sample data register and a sample management processor. The sample encoding processor is used to encode the sample, the sample data register is used to store the sample data, and the sample management processor is used to control and manage the use of the sample.

[0025] The feature selection unit includes an initial screening processor, a statistical analysis processor and a feature evaluation processor. The initial screening processor is used to perform simple screening and filtering of features that are not universal and obviously irrelevant. The statistical analysis processor is used to analyze the correlation between each feature and computing power requirements. The feature evaluation processor is used to evaluate the importance of the feature to the computing power requirements.

[0026] The statistical analysis processor calculates the correlation value r between the feature and the computing power requirement according to the following formula: .

[0027] Where m is the number of samples, x i represents the eigenvalue of the i-th sample, y i represents the computing power requirement of the i-th sample, is the characteristic mean of m samples, is the mean computing power requirement of m samples.

[0028] The feature evaluation processor sorts the features from large to small according to the associated values, and calculates the importance index E of the features according to the following formula: .

[0029] Among them, u is the sorting number of the feature, and n is the number of features.

[0030] The feature engineering unit includes a feature interaction generation processor, a multidimensional feature aggregation processor and an engineering record processor. The feature interaction generation processor is used to interactively process multiple features to generate new deep features. The multidimensional feature aggregation processor is used to select multiple deep features to form a feature vector. The engineering record processor is used to record the interaction method and the vector composition method.

[0031] The feature interaction generation processor interactively processes the features according to the following formula to obtain the deep feature X: .

[0032] Among them, H i () represents the i-th interaction function, x(1) and x(2) represent two interaction features, E(1) represents the importance index of the first interaction feature, and E(2) represents the importance index of the second interaction feature.

[0033] The feature interaction generation processor calculates the weight value Q of the deep feature according to the following formula: .

[0034] The multi-dimensional feature aggregation processor combines deep features so that the sum of weight values ​​of all elements in the feature vector is within the aggregation range, and the combined deep features are sorted from large to small according to the weight values ​​as elements of the feature vector.

[0035] The feature conversion unit includes a type conversion processor and a normalization conversion processor. The type conversion processor is used to perform a uniform conversion process on the data type in the feature vector, and the normalization conversion processor is used to perform a normalization process on the numerical value in the feature vector.

[0036] The model construction unit includes an input configuration processor, a hidden design processor and an output construction processor. The input configuration processor is used to configure the input layer structure of the model to receive and process feature vectors. The hidden design processor is used to construct the number, type and connection method of hidden layer nodes to process feature vectors. The output construction processor is used to construct the output layer structure of the model to output predicted values.

[0037] The learning and training unit includes a sample input processor, a training feedback processor and a training control processor. The sample input processor is used to obtain sample data and place the corresponding feature vector at the correct position of the input layer. The training feedback processor is used to perform feedback statistics on the calculation status of each node in the hidden layer. The training control processor controls the output state of the nodes in the hidden layer based on the statistical data.

[0038] The training feedback processor creates a separate counting storage space for each node. When a node completes a calculation and sends the result to the next node, the count value in the counting storage space corresponding to the node is increased by 1. When the output layer outputs the processing result of a sample, the count values ​​in all counting storage spaces are reduced by 1.

[0039] The training control processor calculates the crowding index P of each node according to the following formula: .

[0040] Among them, a is the range length, N0 is the count value of this node, N iis the count value of the i-th subsequent node.

[0041] When the congestion index is less than the threshold, the node can directly send the calculation result to the next node, otherwise the output channel is closed until the congestion index is less than the threshold.

[0042] The model optimization unit includes a parameter optimization processor and a node pruning processor. The parameter optimization processor is used to optimize the calculation parameters in the nodes, and the node pruning processor is used to prune the nodes that have less influence on the prediction results.

[0043] The model storage unit includes a model extraction processor, a structure information register and a parameter information register. The model extraction processor is used to extract model data from the model training module, the structure information register is used to store the structure information of the model data, and the parameter information register is used to store the parameter information of the model data.

[0044] The real-time prediction unit includes a model loading processor, a real-time receiving processor and a prediction driving processor. The model loading processor is used to load model data, the real-time receiving processor is used to receive prediction applications, and the prediction driving processor is used to drive the model to process the data in the prediction application to obtain a prediction result.

[0045] The feedback adjustment unit includes a result waiting processor, an error calculation processor and a task integration processor. The result waiting processor is used to wait for receiving actual operating resource data of the task in the prediction application, the error calculation processor is used to calculate the deviation between the prediction data and the operating data, and the task integration processor is used to integrate the deviation information with the task information and feed it back to the data collection module as a new sample.

[0046] The i appearing in the above text is an ordinal number used to indicate a sequence number.

[0047] Some of the code for this system is described below: import numpy as np from sklearn.preprocessing import StandardScaler from sklearn.neural_network import MLPRegressor class ComputePowerPredictionSystem: def __init__(self): # Initialize each module self.data_collection = DataCollectionModule() self.feature_extraction = FeatureExtractionModule() self.model_training = ModelTrainingModule() self.prediction_execution = PredictionExecutionModule() def run(self, raw_data): # The overall process of data collection, feature extraction, model training and prediction execution sample_data = self.data_collection.collect_data(raw_data) features = self.feature_extraction.extract_features(sample_data) self.model_training.train_model(features) predictions = self.prediction_execution.predict(features) return predictions class DataCollectionModule: def collect_data(self, raw_data): #Simulation data collection module, including data preprocessing and storage print("Collecting and preprocessing data...") # Example: Standardization scaler = StandardScaler() processed_data = scaler.fit_transform(raw_data) return processed_data class FeatureExtractionModule: def extract_features(self, data): # Feature extraction module, including feature selection, engineering and transformation print("Extracting features...") # Assume simple feature selection and transformation selected_features = data[:, :5] # Select the first 5 hypothesized relevant features return selected_features class ModelTrainingModule: def __init__(self): self.model = MLPRegressor(hidden_layer_sizes=(50,), max_iter=100) def train_model(self, features): # Model training module print("Training model...") X_train, y_train = features[:, :-1], features[:, -1] self.model.fit(X_train, y_train) class PredictionExecutionModule: def __init__(self, model=None): self.model = model def predict(self, features): # Speculative execution module print("Predicting compute power requirement...") predictions = self.model.predict(features) return predictions Now we use this system as the experimental group and the ordinary system as the control group. We train 10 batches of samples and measure the number of samples trained within 100 seconds. The comparison results are as follows: Figure 6 shown.

[0048] The contents disclosed above are only preferred feasible embodiments of the present invention, and do not limit the protection scope of the present invention. Therefore, all equivalent technical changes made using the contents of the present invention specification and drawings are included in the protection scope of the present invention. In addition, the elements therein can be updated as technology develops.

Claims

1. A computing power demand prediction system using deep learning, characterized in that: It includes data collection module, feature extraction module, model training module and prediction execution module; The data collection module is used to collect sample data for training, the feature extraction module is used to extract feature information from the sample data, the model training module is used to perform deep learning training on the sample data, and the prediction execution module predicts the computing power requirements of the task based on the learning results; The data collection module includes a task collection unit, a data preprocessing unit and a sample storage unit. The task collection unit is used to collect the task data after execution. The data preprocessing unit is used to uniformly process the task data into data in the form of samples. The sample storage unit is used to encode and store the samples. The feature extraction unit includes a feature selection unit, a feature engineering unit and a feature conversion unit, wherein the feature selection unit is used to select key features related to computing power requirements, the feature engineering unit is used to construct high-order features from the selected features, and the feature conversion unit is used to convert the constructed features into a format supported by the model; The model training module includes a model construction unit, a learning and training unit and a model optimization unit. The model construction unit is used to construct a neural network structure to generate a training model. The learning and training unit is used to control the training model to perform learning and training on sample data. The model optimization unit optimizes the training model based on the training results. The prediction execution module includes a model storage unit, a real-time prediction unit and a feedback adjustment unit. The model storage unit is used to store model data used for prediction. The real-time prediction unit is used to receive prediction tasks and use the model to make predictions. The feedback adjustment unit is used to collect actual running data of the task, process the error information obtained, and then feed it back to the data collection module.

2. A computing power demand prediction system using deep learning as claimed in claim 1, characterized in that: The feature selection unit includes an initial screening processor, a statistical analysis processor and a feature evaluation processor, wherein the initial screening processor is used to perform simple screening and filtering of features that are not universal and obviously irrelevant, the statistical analysis processor is used to analyze the correlation between each feature and the computing power requirement, and the feature evaluation processor is used to evaluate the importance of the feature to the computing power requirement; The feature evaluation processor sorts the features from large to small according to the associated values, and calculates the importance index E of the features according to the following formula: ； Among them, u is the sorting number of the feature, n is the number of features, and r is the associated value.

3. A computing power demand prediction system using deep learning as claimed in claim 2, characterized in that: The feature engineering unit includes a feature interaction generation processor, a multidimensional feature aggregation processor and an engineering record processor, wherein the feature interaction generation processor is used to interactively process multiple features to generate new deep features, the multidimensional feature aggregation processor is used to select multiple deep features to form a feature vector, and the engineering record processor is used to record the interaction method and the vector formation method; The feature interaction generation processor interactively processes the features according to the following formula to obtain the deep feature X: ； Among them, H i () represents the i-th interaction function, x(1) and x(2) represent two interaction features, E(1) represents the importance index of the first interaction feature, and E(2) represents the importance index of the second interaction feature.

4. A computing power demand prediction system using deep learning as claimed in claim 3, characterized in that: The learning and training unit includes a sample input processor, a training feedback processor and a training control processor. The sample input processor is used to obtain sample data and place the corresponding feature vector at the correct position of the input layer. The training feedback processor is used to perform feedback statistics on the calculation status of each node in the hidden layer. The training control processor controls the output state of the nodes in the hidden layer based on the statistical data.

5. A computing power demand prediction system using deep learning as claimed in claim 4, characterized in that: The training control processor calculates the crowding index P of each node according to the following formula: ； Among them, a is the range length, N0 is the count value of this node, N i is the count value of the i-th subsequent node; When the congestion index is less than the threshold, the node can directly send the calculation result to the next node, otherwise the output channel is closed until the congestion index is less than the threshold.

Images