Energy Consumption Optimization Method and System for Smart Computing Centers Based on the Internet of Things

Abstract

This application relates to the field of data processing technology and discloses a method and system for optimizing energy consumption in an intelligent computing center based on the Internet of Things (IoT). The method includes: collecting energy consumption parameters of storage media and data block access behavior parameters to generate a raw dataset; calculating weighted access feature values ​​and the probability of hot / cold state transitions; screening candidate data blocks and calculating energy efficiency evaluation indicators; matching renewable energy supply status to construct a migration task sequence; executing migration operations; and dynamically adjusting prediction parameters. This application solves the problems in existing technologies such as the inability to intelligently predict data hot / cold state transitions, the lack of quantitative assessment of energy efficiency, insufficient coordination between data migration and renewable energy supply, and the inability to adaptively adjust prediction model parameters.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a method and system for optimizing energy consumption in an intelligent computing center based on the Internet of Things. Background Technology

[0002] With the rapid development of IoT technology and the continuous expansion of data center scale, energy management in intelligent computing centers has gradually become one of the core issues in data center operation. Existing IoT-based energy optimization methods for intelligent computing centers mainly involve deploying IoT devices such as temperature sensors and power consumption sensors in the data center to collect environmental parameters and equipment operation data. This data is then used to dynamically adjust the cooling and power supply systems to reduce overall energy consumption. Some technical solutions employ a strategy based on fixed thresholds for tiered data storage management. For example, data that has not been accessed for more than 30 days is classified as cold data and migrated to low-power storage media. Simultaneously, renewable energy sources such as photovoltaic power generation systems are used to power the data center to reduce reliance on the traditional power grid.

[0003] However, existing technologies have several shortcomings. First, methods based on fixed time thresholds to determine data hot / cold status cannot adapt to dynamic changes in data access patterns. For example, the access frequency of financial data surges during month-end settlements; fixed threshold methods cannot predict this periodic change, leading to misclassification of data about to become hot as cold data and migration to low-speed storage media, resulting in a sharp increase in access latency. Second, existing energy optimization methods fail to fully consider the energy cost of the migration operation itself when making data migration decisions, focusing only on energy savings after migration while ignoring the energy consumed in the reading, transmission, and writing processes, resulting in net energy losses for migrating small data blocks or short-term cold data. Third, there is a lack of effective coordination mechanisms between data migration scheduling and the availability of renewable energy. Migration tasks are often executed during peak grid periods, increasing electricity purchase costs, while during periods of sufficient photovoltaic power generation, green energy is not fully utilized for data reorganization and migration operations. Summary of the Invention

[0004] This application provides an IoT-based intelligent computing center energy consumption optimization method and system, which solves the problems in the prior art such as the inability to intelligently predict the transition between hot and cold data states, the lack of quantitative assessment of energy consumption benefits, insufficient coordination between data migration and renewable energy supply, and the inability to adaptively adjust prediction model parameters. It achieves a significant reduction in energy consumption of data center storage systems and a substantial increase in the utilization rate of renewable energy.

[0005] Firstly, this application provides an energy consumption optimization method for intelligent computing centers based on the Internet of Things (IoT), the method comprising:

[0006] Step S1: Collect energy consumption characteristic parameters of different types of storage media in the data center and data block access behavior parameters through the Internet of Things sensor network to generate an original dataset containing time information, data block identifiers and energy consumption characteristics.

[0007] Step S2: Obtain access feature values ​​by weighting the access behavior parameters and time decay factors in the original dataset, and calculate the cold / hot state transition probability of the data block based on the access feature values;

[0008] Step S3: Filter candidate data blocks whose cold / hot state transition probability meets the set conditions, calculate the comparison value between the energy consumption difference of the candidate data block between the current storage medium and the target storage medium and the energy consumption cost of the migration operation, and obtain the energy consumption benefit evaluation index.

[0009] Step S4: Match the energy consumption efficiency assessment indicators with the renewable energy supply status, and select high-efficiency data blocks to construct a migration task sequence during periods of sufficient energy supply.

[0010] Step S5: Execute the media transfer operation of the data block in the migration task sequence, and collect the deviation information between the actual access parameters and the predicted access parameters of the data block after migration for dynamic adjustment of the access feature value calculation parameters.

[0011] Secondly, this application provides an IoT-based intelligent computing center energy consumption optimization system, the IoT-based intelligent computing center energy consumption optimization system comprising:

[0012] The generation module is used to collect energy consumption characteristic parameters of different types of storage media in the data center and data block access behavior parameters through the Internet of Things sensor network, and generate a raw dataset containing time information, data block identifiers and energy consumption characteristics.

[0013] The weighting module is used to perform weighted calculations on the access behavior parameters and time decay factors in the original dataset to obtain access feature values, and to calculate the cold and hot state transition probability of the data block based on the access feature values;

[0014] The calculation module is used to filter candidate data blocks whose cold and hot state transition probabilities meet the set conditions, calculate the energy consumption difference between the candidate data blocks and the target storage medium and the energy consumption cost of the migration operation, and obtain the energy consumption benefit evaluation index.

[0015] The matching module is used to match and judge the energy consumption efficiency assessment indicators with the renewable energy supply status, and select high-efficiency data blocks to construct a migration task sequence during periods of sufficient energy supply.

[0016] The acquisition module is used to perform media transfer operations on data blocks in the migration task sequence, and to acquire the deviation information between the actual access parameters and the predicted access parameters of the data blocks after migration for dynamic adjustment of access feature value calculation parameters.

[0017] Thirdly, an IoT-based intelligent computing center energy consumption optimization device is provided, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor invokes the instructions in the memory to cause the IoT-based intelligent computing center energy consumption optimization device to execute the aforementioned IoT-based intelligent computing center energy consumption optimization method.

[0018] Fourthly, a computer-readable storage medium is provided, wherein instructions are stored therein, which, when executed on a computer, cause the computer to perform the above-described energy consumption optimization method for an IoT-based intelligent computing center.

[0019] The technical solution provided in this application collects energy consumption characteristic parameters of different types of storage media and access behavior parameters of data blocks in a data center through an Internet of Things (IoT) sensor network, and generates a raw dataset. This breaks through the limitation of existing technologies that only monitor total energy consumption at a macro level, and realizes a refined correlation between access behavior at the data block level and the corresponding storage media's energy consumption characteristics, laying a data foundation for subsequent intelligent decision-making. By weighting access behavior parameters with a time decay factor to obtain access characteristic values ​​and calculating the probability of data block hot / cold state transitions, this application introduces the dynamic impact of the time forgetting effect on data activity. Compared with the static method in existing technologies that determines hot / cold states based on fixed time thresholds, this method can accurately capture the natural decay law and periodic change characteristics of data access patterns over time, predict in advance when a data block is about to transition from a hot state to a cold state, and avoid the problem of increased access latency caused by misjudging data that is about to become hot as cold data. This application derives an energy efficiency evaluation index by screening candidate data blocks whose cold / hot state transition probabilities meet set conditions and calculating the energy consumption difference compared to the energy consumption cost of migration operations. It establishes a quantitative evaluation model that simultaneously considers energy savings and migration energy costs, addressing the problem in existing technologies that only focus on post-migration energy savings while ignoring net energy losses caused by energy consumption during the migration process. By dividing the net energy savings by the access latency, a joint trade-off between energy optimization and service quality assurance is achieved. Furthermore, this application matches the energy efficiency evaluation index with the renewable energy supply status and selects high-efficiency data blocks to construct a migration task sequence during periods of sufficient energy supply. This enables intelligent coordination between data migration scheduling and renewable energy supply such as photovoltaic power generation, maximizing the use of green energy windows to execute energy-intensive migration tasks, significantly reducing grid power purchase costs and improving renewable energy utilization. This application establishes a complete closed-loop optimization mechanism from prediction to execution to feedback calibration by performing media transfer operations on data blocks in the migration task sequence and collecting the deviation information between the actual access parameters and the predicted access parameters of the migrated data blocks for dynamic adjustment of access feature value calculation parameters. This solves the problem in the prior art where static configuration of prediction model parameters leads to a decrease in accuracy as business scenarios evolve, and enables the entire energy consumption optimization system to have the ability to learn adaptively and continuously improve.

[0020] Bidirectional Long Short-Term Memory (LSTM) networks process the historical access time-series features of data blocks simultaneously in both forward and backward directions. This allows them to learn the decay trend of data activity from past access records and to understand the inherent patterns of current access behavior using information from subsequent time points in the sequence. This bidirectional time-series modeling capability enables the network to accurately identify periodic fluctuations and sudden changes in data access patterns, exhibiting higher prediction accuracy and adaptability to complex access patterns compared to unidirectional recurrent neural networks or simple statistical prediction methods. The time decay factor introduced in this application quantifies the weakening effect of data block idle time on its activity through an exponential decay function. The weighted combination of this decay factor and access frequency allows the weighted access feature value to simultaneously reflect the historical access intensity and the impact of time forgetting. This weighting mechanism is particularly important in data center storage management scenarios because the hot / cold status of data depends not only on past access frequency but also on whether it is still frequently accessed recently. The weighted feature value, as input to the neural network, significantly improves the accuracy of predicting the probability of hot / cold status transitions. The energy efficiency evaluation index constructed in this application achieves a mathematical unification of energy efficiency optimization objectives and service quality constraints at the algorithm level by calculating the difference between energy savings and migration operation energy costs, and then dividing by the access latency. This index serves as a ranking criterion in data migration decisions, ensuring that data blocks that achieve the maximum energy savings at an acceptable latency cost are migrated first. This quantitative index design for multi-objective optimization makes the algorithm's decisions clearly interpretable and operable. The incremental learning mechanism designed in this application updates the parameters of the bidirectional long short-term memory network using a small learning rate. In its algorithmic implementation, this mechanism avoids catastrophic forgetting of the network's learned knowledge by controlling the step size of gradient descent, enabling the network to gradually adapt to new business scenario changes while retaining its understanding of historical access patterns. This online learning capability is crucial for long-running applications like data centers with constantly evolving business models, ensuring the long-term effectiveness and robustness of the energy efficiency optimization algorithm. Attached Figure Description

[0021] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a schematic diagram of one embodiment of the IoT-based intelligent computing center energy consumption optimization method in this application.

[0023] Figure 2 This is a schematic diagram comparing the energy consumption optimization effects of IoT-based intelligent computing centers in the embodiments of this application;

[0024] Figure 3 This is a schematic diagram of one embodiment of the IoT-based intelligent computing center energy consumption optimization system in this application.

[0025] Figure 4 This is a schematic block diagram of the energy consumption optimization device for an intelligent computing center based on the Internet of Things in this embodiment of the invention. Detailed Implementation

[0026] This application provides an energy consumption optimization method and system for an IoT-based intelligent computing center. The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or device that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0027] For ease of understanding, the specific process of the embodiments of this application is described below. Please refer to [link / reference]. Figure 1 One embodiment of the IoT-based intelligent computing center energy consumption optimization method in this application includes:

[0028] Step S1: Collect energy consumption characteristic parameters of different types of storage media in the data center and data block access behavior parameters through the Internet of Things sensor network to generate an original dataset containing time information, data block identifiers and energy consumption characteristics.

[0029] Step S2: Obtain access feature values ​​by weighting the access behavior parameters and time decay factors in the original dataset, and calculate the cold / hot state transition probability of the data block based on the access feature values.

[0030] Step S3: Select candidate data blocks whose cold / hot state transition probability meets the set conditions, calculate the energy consumption difference between the current storage medium and the target storage medium of the candidate data blocks and the energy consumption cost of the migration operation, and obtain the energy consumption benefit evaluation index.

[0031] Step S4: Match the energy consumption efficiency assessment indicators with the renewable energy supply status, and select high-efficiency data blocks to construct a migration task sequence during periods of sufficient energy supply.

[0032] Step S5: Perform media transfer operations on the data blocks in the migration task sequence, and collect the deviation information between the actual access parameters and the predicted access parameters of the migrated data blocks for dynamic adjustment of the access feature value calculation parameters.

[0033] It is understood that the executing entity of this application can be an IoT-based intelligent computing center energy consumption optimization system, or it can be a terminal or a server; the specific implementation is not limited here. This application's embodiment uses a server as an example for illustration.

[0034] Specifically, a dedicated power consumption monitoring module is deployed within the data center storage rack to accurately collect the energy consumption characteristics of heterogeneous storage media. This module is directly connected to the power supply circuits of disk arrays, solid-state drives, and Blu-ray storage devices, measuring the instantaneous power consumption of each storage medium in real time and calculating the unit power consumption value based on the storage capacity, thus forming a storage medium energy consumption parameter set. Simultaneously, a data access behavior monitoring sensor is deployed. This sensor is embedded in the storage system's I / O controller, intercepting and recording metadata information for each data read / write operation, including the unique identifier of the accessed data block, the precise timestamp of the access, the time interval between two consecutive accesses, and the storage size of the data block. This information constitutes the access behavior parameter set. The system then performs an association matching operation, using the data block identifier as the key, to match the storage medium type and power consumption data in the energy consumption parameter set with the access characteristics in the access behavior parameter set, generating associated data records. Each record fully describes the access status and energy consumption status of a specific data block at a specific moment. Normalization is performed on the key parameter of access frequency. The maximum access frequency among all data blocks is found as the normalization benchmark. The access frequency of each data block is divided by this maximum value, mapping it to a numerical range of 0 to 1. This eliminates the dimensional impact of fluctuations in access volume over different time periods, ultimately generating a structured original dataset. For example, a data center collects data block A stored on a disk array within a week. The unit power consumption is 15W / TB, the access frequency is 80 times / hour, the access interval is 45 seconds, and the data block size is 500MB. These parameters, after correlation and normalization, form a complete record in the original dataset.

[0035] The access time intervals of data blocks are extracted one by one from the original dataset. This time interval represents the elapsed time from the last access to the current moment. The time interval value is multiplied by a preset decay coefficient, and the result is negative. Then, an exponential operation is performed with the natural constant as the base to obtain a time decay factor between 0 and 1. This factor reflects the natural decay of data block activity over time. The normalized access frequencies in the original dataset are then multiplied element-wise with the corresponding time decay factors to obtain weighted access feature values. This value comprehensively considers the absolute magnitude of the access frequency and the time decay effect, more accurately characterizing the current actual activity level of the data blocks. Weighted access feature values ​​within continuous time windows are extracted from the original dataset in chronological order to form a time-series feature sequence. The sequence length is set to 168 time points, i.e., 7 days of hourly data. The temporal feature sequence is input into a bidirectional Long Short-Term Memory (LSTM) network, which contains two independent LSTM layers: a forward layer that processes data from the beginning to the end of the sequence to capture historical access trends, and a backward layer that processes data from the end to the beginning of the sequence to identify possible future access patterns. The hidden states from both directions are concatenated at each time step to form a more comprehensive feature representation. After multi-layer processing, a fixed-dimensional hidden state vector is obtained. This hidden state vector is then input into an activation function layer, which uses the sigmoid function to map the vector to a probability range of 0 to 1. The output value represents the probability of a data block transitioning from a hot to a cold state within the next 72 hours. For example, if data block B has an access interval of 72 hours (3 days) and has not been accessed, with a preset decay coefficient of 0.035, the calculated time decay factor is 0.114, and its normalized access frequency is 0.2. Multiplying the two together yields a weighted access feature value of 0.023. The 168 weighted access feature values ​​of this data block over the past 7 days form a time-series feature sequence. After inputting this sequence into a bidirectional long short-term memory network, the forward layer identifies a trend of continuously decreasing access frequency, and the backward layer predicts that the probability of future access is extremely low. The combined output of both results in a cold / hot state transition probability of 0.86.

[0036] The cold / hot state transition probability of each data block calculated by S2 is compared one by one with a preset probability threshold, which is set to 0.75. Data blocks with a cold / hot state transition probability greater than this threshold are marked as candidate data blocks, and these data blocks are judged to be highly likely to be converted to cold data. The unit power consumption value of the current storage medium and the unit power consumption value of the target storage medium to which the candidate data block is planned to be migrated are retrieved from the original dataset. The difference between the two power consumption values ​​represents the power saving space per unit storage capacity. This power consumption difference is multiplied by the expected duration of the data block in the cold state. The duration is additionally predicted by the output layer of the bidirectional long short-term memory network, which is usually 720 hours or 30 days. The product result is the energy saving of the candidate data block, in watt-hours. The energy cost of migrating candidate data blocks consists of three components: read energy cost (calculated by multiplying read power by read time, where read time equals data block size divided by disk read speed), network transmission energy cost (calculated by multiplying network device power consumption by transmission time, where transmission time equals data block size divided by network bandwidth), and write energy cost (calculated by multiplying write power by write time, where write time equals data block size divided by target medium write speed). The sum of these three components yields the total energy cost of the migration operation. Subtracting the migration operation energy cost from the energy savings yields the net energy benefit. Dividing this net energy benefit by the access latency after the data block is migrated to the target medium provides an energy efficiency evaluation index per unit latency time. This index simultaneously measures the energy saving effect and the impact on service quality. For example, candidate data block C has a cold / hot state transition probability of 0.83, which is greater than 0.75, and is therefore selected. Currently stored on the disk array, its power consumption is 15W / TB, while the target Blu-ray storage has a power consumption of 0.5W / TB, resulting in a power difference of 14.5W / TB. The data block size is 1GB (0.001TB), and the expected cold state retention time is 720 hours. The energy saving is 14.5 × 0.001 × 720 = 10.44Wh. The read power is 25W, and the read speed is 100... MB / s read time 10 seconds, read energy consumption 0.069Wh; network power 15W, bandwidth 10Gbps, transmission time 0.8 seconds, transmission energy consumption 0.003Wh; write power 30W, write speed 80MB / s, write time 12.5 seconds, write energy consumption 0.104Wh; total migration energy consumption 0.176Wh; net energy saving benefit 10.264Wh; Blu-ray storage access latency 5 seconds; energy saving benefit assessment index is 10.264 ÷ 5 = 2.053Wh / s.

[0037] The system collects real-time power generation data from the rooftop photovoltaic (PV) system and the remaining capacity data from the basement battery storage system. PV power generation reflects the instantaneous power generation capacity under current solar radiation intensity, while the remaining battery capacity represents the currently available stored electricity. Multiplying the remaining battery capacity by the discharge efficiency of 0.95 yields the actual usable electricity, which is then added to the PV power generation to obtain the renewable energy quota. This quota represents the total amount of green energy available for data migration at the current moment. The renewable energy quota value is compared with a preset high energy threshold of 50 kWh. When the renewable energy quota exceeds this threshold, the current period is considered a period of sufficient energy supply, suitable for performing energy-intensive data migration tasks to maximize the use of renewable energy and reduce grid electricity purchase costs. During periods of sufficient energy supply, candidate data blocks are sorted in descending order according to energy efficiency evaluation indicators, with the data blocks having the highest energy efficiency evaluation indicator values ​​placed at the top. These data blocks offer the highest energy savings per unit of latency. The ratio of renewable energy quota to the average migration energy consumption of a single data block is used to determine the upper limit of the number of migrated data blocks. The average migration energy consumption is calculated by statistically analyzing the energy consumption of historical migration tasks, and is typically 0.2Wh. The ratio is rounded down to obtain the maximum number of data blocks that can be migrated under the current energy constraints. Data blocks are selected sequentially from the sorted candidate data block sequence, with the number of selections not exceeding the upper limit of the number of migrated data blocks. A migration task sequence is constructed, in which each element contains a data block identifier, source storage medium type, target storage medium type, and migration priority determined based on energy efficiency evaluation indicators. For example, if the current photovoltaic power generation is 40kW, the remaining battery capacity is 15kWh, the discharge efficiency is 0.95, the actual usable electricity is 14.25kWh, and the renewable energy quota is 40 + 14.25 = 54.25kWh. Since the energy supply exceeds 50kWh, it is considered a period of sufficient energy supply. After sorting, the top three energy efficiency evaluation indicators for the candidate data blocks are 2.053, 1.876, and 1.642Wh / s, respectively. The renewable energy quota of 54.25kWh is 54250Wh, the average migration energy consumption is 0.2Wh, and the maximum number of data blocks that can be migrated is 54250 ÷ 0.2 = 271250. The top 271250 data blocks are selected from the sorted queue to construct a migration task sequence. The sequence includes data block C, identified as source media disk and target media Blu-ray priority 1.

[0038] Based on the data block identifiers and target storage media information recorded in the migration task sequence, media transfer operations are performed sequentially in descending order of migration priority. For each data block, a read operation is performed to read the data block content from the source storage media into a memory buffer. A transfer operation transmits the data from the source storage node to the target storage node via the data center's internal network. A write operation writes the data to the target storage media, completing the physical migration. After the media transfer is completed, a continuous monitoring mechanism is initiated. Access events of the migrated data blocks are recorded hourly for a preset monitoring period of 30 days. The total number of accesses during this period is divided by the monitoring duration to obtain the actual access frequency, which reflects the true access activity of the data blocks after migration. The difference between the actual access frequency and the predicted access frequency obtained before migration using a bidirectional long short-term memory network based on weighted access feature values ​​is calculated. The absolute value of the difference is divided by the predicted access frequency to obtain the relative deviation value, which quantifies the prediction accuracy. When the relative deviation exceeds the preset deviation threshold of 0.3, it indicates a large prediction error. The weighted access feature value sequence of the data block for the consecutive 168 hours before migration and the actual recorded access frequency after migration are marked as correction samples. The input feature of the correction samples is the historical weighted access feature value sequence, and the true label is whether the actual access frequency is lower than the cold data criterion. The accumulated correction samples are batch-input into a bidirectional long short-term memory network, with a learning rate set to one-tenth of the initial training learning rate of 0.001, i.e., 0.0001. The gradient of the prediction error with respect to the network weight parameters is calculated using the backpropagation algorithm, and the weight parameters are updated along the gradient descent direction. Incremental learning is completed after 10 iterations, enabling the network to learn from the prediction error and improve future prediction accuracy. All data migration tasks executed in the past week are statistically analyzed. The ratio of the actual migration energy consumption to the predicted migration energy consumption for each task is calculated. The average of all ratios is used to obtain the energy consumption calibration coefficient. When the deviation of the energy consumption calibration coefficient from the preset benchmark value 1 exceeds the preset calibration threshold of 0.15, the preset attenuation coefficient is multiplied by the energy consumption calibration coefficient to obtain the updated attenuation coefficient value. The updated attenuation coefficient is used to calculate the time attenuation factor in the subsequent S2 step. For example, after data block D was migrated to Blu-ray storage, the actual access frequency over 30 days was 0.05 times / hour, and the predicted access frequency was 0.15 times / hour. The relative deviation was |0.05-0.15|÷0.15=0.67, which exceeded 0.3. The 168-hour weighted access feature value sequence of this data block and the actual access frequency were used to construct a correction sample input to a bidirectional long short-term memory network. The learning rate was set to 0.0001, and 10 rounds of incremental learning were performed to update the network parameters. The average ratio of actual energy consumption to predicted energy consumption for 20 migration tasks in the past week was 1.18. The deviation of the energy consumption calibration coefficient 1.18 from the baseline value 1 was 0.18, which exceeded 0.15. The original preset attenuation coefficient 0.035 was multiplied by 1.18 to obtain the updated attenuation coefficient 0.0413.

[0039] In one specific embodiment, step S1 includes:

[0040] Deploy power consumption monitoring modules in the storage racks of the data center to collect the unit power consumption values ​​of disk arrays, solid-state drives and Blu-ray storage media to obtain a set of energy consumption parameters for storage media;

[0041] By using data access behavior monitoring sensors to record the access frequency, access time interval, and data block size of each data block at the timestamp, a set of access behavior parameters can be obtained.

[0042] The storage medium energy consumption parameter set and the access behavior parameter set are associated and matched according to the data block identifier to generate associated data records containing timestamp, data block identifier, access frequency, current storage medium type and real-time power consumption value;

[0043] The access frequency in the associated data records is normalized by dividing the access frequency by the maximum access frequency value and mapping it to a preset numerical range to obtain the original dataset.

[0044] Specifically, the power consumption monitoring module is deployed in the power distribution unit inside the storage rack. It monitors the current and voltage values ​​flowing through the disk array, solid-state drive, and Blu-ray storage device in real time using current and voltage sensors. The instantaneous power is obtained by multiplying the current and voltage, and then divided by the total capacity of the corresponding storage device to obtain the unit power consumption value. The unit power consumption value represents the power consumed per unit of storage capacity. Different types of storage media have significantly different unit power consumption values ​​due to differences in their physical structure and working principle. The disk array requires higher power consumption due to the rotation of mechanical parts and the seek of the read / write head. Solid-state drives use flash memory chips with moderate power consumption. Blu-ray storage has extremely low power consumption when storing data statically. The type identification of the three types of storage media and their corresponding unit power consumption values ​​are stored in a data table to form a storage media energy consumption parameter set. This parameter set serves as the basic data source for subsequent energy consumption calculations. Data access behavior monitoring sensors are embedded in the block device driver layer of the storage system. This layer, located in the operating system kernel space, is responsible for managing all block device input / output operations. The sensors intercept all I / O requests sent to the storage device through hook functions. The I / O requests contain metadata information such as request type, target address, and data length. The logical block address of the target data block is parsed from the request header as a unique identifier for the data block. The system timestamp of the request arriving at the storage subsystem is recorded. The system timestamp is accurate to the millisecond level, reflecting the exact moment the access occurred. By querying the timestamp of the last access to the data block in the historical access record table, the current timestamp and the last access time are calculated. The difference between the timestamps yields the access time interval, which reflects the access activity of the data block. The shorter the interval, the more frequent the access. The total number of times the data block is accessed within a unit time window yields the access frequency, which directly quantifies the popularity of the data block. The data transfer length is extracted from the parameter field of the I / O request to obtain the data block size. The data block size affects the energy and time costs of the migration operation. The data block identifier, timestamp, access frequency, access time interval, and data block size are written to the log file according to a predefined data structure. The log file uses an append-only writing mode to ensure that no access events are lost. All log records are summarized to form an access behavior parameter set.The association matching operation initiates a dedicated data processing thread. This thread periodically reads new records from the access behavior parameter set, extracts the data block identifier field from each record, and queries the metadata management table maintained by the storage system based on the data block identifier. The metadata management table records the physical location of each data block in terms of the type of storage medium it is currently stored in. It retrieves the storage medium type identifier corresponding to the data block, and then uses the storage medium type identifier as the key to look up the corresponding unit power consumption value in the storage medium energy consumption parameter set. The timestamp, data block identifier, access frequency, access time interval, and data block size fields from the access behavior parameter set are horizontally concatenated with the current storage medium type and unit power consumption value fields obtained from the storage medium energy consumption parameter set. The concatenation operation constructs a new data structure in memory. This data structure contains complete information describing the data block access behavior and energy consumption characteristics. The generated associated data records are written to a dedicated table in the relational database. The database table is indexed by the timestamp field to accelerate subsequent query operations. All associated data records constitute a comprehensive dataset that simultaneously includes the access behavior dimension and the energy consumption characteristic dimension. The normalization module reads the associated data record table. The access frequencies of different data blocks in this table vary by orders of magnitude due to differences in business characteristics. Directly using the raw access frequencies would result in an excessively large numerical range, affecting the convergence speed and prediction accuracy of subsequent machine learning algorithms. The normalization process iterates through the access frequency field of all records in the table, finding the global maximum access frequency through a single full scan. This maximum value represents the access intensity of the hottest data block within the observation time window and serves as the baseline value for normalization calculations. Normalization is then performed on each record in the table, reading the raw access frequency value for that record. Dividing the raw value by the baseline value yields the global maximum access frequency. This division operation maps all access frequencies to zero. The normalized access frequency range is defined as follows: zero indicates no access within the time window, one indicates the access frequency has reached its maximum value, and intermediate values ​​reflect the relative access intensity proportionally. The normalized access frequency value is used to replace the access frequency field of the record in the database table. The database update operation is then performed to persist the normalized result. The associated data record table after all records have been updated is the original dataset. Each record in the original dataset contains a timestamp in the time dimension, a data block identifier in the identifier dimension, normalized access behavior characteristics, and corresponding energy consumption characteristics. The normalized access frequency in the dataset eliminates the influence of units, making the data block access patterns under different time periods and business scenarios comparable.

[0045] In one specific embodiment, step S2 includes:

[0046] The access time interval of each data block is extracted from the original dataset. The access time interval is then subjected to a negative exponential operation with the preset decay coefficient, and the result is taken as a power of the natural constant to obtain the time decay factor of each data block.

[0047] The normalized access frequency in the original dataset is multiplied by the corresponding time decay factor to obtain a weighted access feature value that reflects the actual activity level of the data block.

[0048] Weighted access feature values ​​within continuous time windows in the original dataset are extracted to construct a temporal feature sequence. The temporal feature sequence is then input into a bidirectional long short-term memory network for forward and backward temporal feature extraction to obtain the hidden state vector of data block access trends.

[0049] The hidden state vector is nonlinearly mapped through an activation function layer, and the output data block is transformed from a hot state to a cold state within a future time window, with the probability of the transformation from a hot state to a cold state.

[0050] Specifically, the access time interval field of each data block is read one by one from the associated data record table of the original dataset. The access time interval represents the elapsed time between the current time and the last time the data block was accessed. The longer the time interval, the longer the data block has been idle and the lower its activity. The access time interval value is multiplied by a preset decay coefficient. The preset decay coefficient is a parameter that controls the decay rate. The larger the value, the faster the activity decays over time. The product result is negative, and this negative number is used as the exponent. The power operation is performed with the natural constant as the base. The mathematical property of the power operation makes the calculation result decrease exponentially as the access time interval increases. Finally, a time decay factor between zero and one is obtained. The time decay factor quantifies the weakening effect of the passage of time on data activity. The time interval of a newly accessed data block is close to zero and the time decay factor is close to one. The time decay factor of a data block that has not been accessed for a long time is close to zero. Read the normalized access frequency field and the corresponding time decay factor of each record in the original dataset, and multiply the two values. The normalized access frequency reflects the intensity of historical access, and the time decay factor reflects the time forgetting effect. Multiplying the two values ​​yields the weighted access feature value. The weighted access feature value takes into account both the absolute magnitude of the access frequency and the influence of time decay. Compared with simply using the access frequency, it more accurately reflects the current actual activity level of the data block. Data blocks with high access frequency but no access for a long time have a lower weighted value, while data blocks with low access frequency but frequent recent access have a higher weighted value. Write the calculated weighted access feature value back to the newly added field of the corresponding record in the database table.Weighted access feature value sequences of specific data blocks within a continuous time window are extracted from the original dataset in timestamp order. The time window length is set to a fixed number of time points, and each time point within the window corresponds to a weighted access feature value. These are arranged chronologically to form a one-dimensional numerical sequence, the time-series feature sequence. This time-series feature sequence preserves the complete trajectory of the data block access pattern's evolution over time. The time-series feature sequence is then fed as input into a bidirectional long short-term memory (LSTM) network. The bidirectional LSM network is a deep learning model specifically designed for processing time-series data, containing two independent recurrent neural network branches: a forward LSM layer and a backward LSM layer. The forward LSM layer processes the input data sequentially from the starting time point of the time-series feature sequence, combining the current weighted access feature value with the hidden state from the previous time step. A forget gate determines how much historical information to retain. The input gate determines how much current information to receive, and the output gate determines how much information to pass to the next time step. After processing all time points in a loop, the historical evolution trend of data block access patterns is captured. The backward long short-term memory layer processes the input data one by one in reverse chronological order, starting from the end time point of the temporal feature sequence. The reverse processing mechanism enables the network to understand the access behavior at the current time using information from future time points. After processing all time points in a loop, the future change trend of access patterns is identified. The hidden state vectors of the forward long short-term memory layer and the backward long short-term memory layer at each time step are concatenated. The concatenated vector contains both past trend information and future prediction information, forming a more comprehensive feature expression. The concatenated vector corresponding to the last time point of the temporal feature sequence is taken as the hidden state vector of the entire sequence. The hidden state vector is a high-dimensional numerical vector containing the abstract features learned by the network from the temporal data. The hidden state vector is input into the activation function layer, which contains a fully connected layer and a Sigmoid activation function. The fully connected layer maps the high-dimensional hidden state vector into a single value through a weight matrix. The Sigmoid activation function performs a non-linear transformation on this value. The mathematical form of the Sigmoid function compresses the output result into the range of zero to one, regardless of how large or small the input value is. The output value is interpreted as the probability that the data block will change from a hot state to a cold state within a preset time window in the future. A value close to one indicates a very high probability of changing to cold data, a value close to zero indicates a very high probability of remaining in the hot data state, and a value in the middle indicates that the hot and cold state is uncertain. This probability value is recorded in a database table as a field for the probability of hot and cold state transition.

[0051] In one specific embodiment, step S3 includes:

[0052] The probability of switching between hot and cold states is compared with a preset probability threshold, and data blocks whose probability of switching between hot and cold states is greater than the preset probability threshold are selected as candidate data blocks.

[0053] Extract the current unit power consumption value of the storage medium and the unit power consumption value of the target storage medium of the candidate data block from the original dataset, calculate the power consumption difference between the two and the product of the expected cold state holding time of the data block, and obtain the energy saving of the candidate data block.

[0054] Calculate the sum of the energy consumption for reading, network transmission, and writing of the candidate data block to obtain the energy cost of the migration operation from the current storage medium to the target storage medium;

[0055] The energy efficiency evaluation index for candidate data blocks is obtained by calculating the difference between the energy savings and the energy cost of the migration operation and dividing it by the data block access latency.

[0056] Specifically, the cold / hot state transition probability field of all data blocks is read from the database table. The cold / hot state transition probability value of each data block is compared with a preset probability threshold, which is the critical value for determining whether a data block is about to become cold data. Data blocks with a cold / hot state transition probability greater than the preset probability threshold are selected and considered to have a high probability of becoming cold data. The selected data block identifiers are recorded in a candidate data block list, which contains all data blocks worth considering migrating to low-power storage media. For each data block in the candidate data block list, the current storage media type field is queried from the associated data record table of the original dataset. Based on the storage media type, the corresponding unit power consumption value is retrieved from the storage media energy consumption parameter set as the current storage media unit power consumption value. Simultaneously, the target storage media type to which the data block is planned to migrate is determined. The target storage media is typically selected as one with a significantly lower unit power consumption than the current media. The unit power consumption value corresponding to the target storage media type is retrieved from the storage media energy consumption parameter set as the target storage media unit power consumption value. The power consumption difference is calculated by subtracting the target storage media unit power consumption value from the current storage media unit power consumption value. The power consumption difference represents the power saved per unit time after the migration of a unit of storage capacity. The data block size field of the data block is extracted from the original dataset, and the data block size is converted from bytes to TB to obtain the storage capacity. The expected cold state duration of the data block is obtained from the auxiliary output of the bidirectional long short-term memory network as the expected cold state duration. This duration is the duration for which the network predicts the data block will remain in a low access frequency state based on historical access patterns. The power consumption difference is multiplied by the storage capacity and then by the expected cold state duration to obtain the energy saving amount. The energy saving amount represents the total energy saved during the entire cold state if the migration operation is performed. The energy cost of migrating candidate data blocks comprises three energy components: read energy cost (calculated by multiplying read power by read time, where read power is the power consumed by the current storage medium during a read operation, and read time equals the data block size divided by the read speed of the current storage medium, where read speed is a performance parameter representing the amount of data that can be read per unit time), network transmission energy cost (calculated by multiplying network device power consumption by transmission time, where network device power consumption is the operating power of network switches and routers within the data center, and transmission time equals the data block size divided by network bandwidth, where network bandwidth is the transmission rate of the network within the data center), and write energy cost (calculated by multiplying write power by write time, where write power is the power consumed by the target storage medium during a write operation, and write time equals the data block size divided by the write speed of the target storage medium, where write speed is a performance parameter). The energy cost of the migration operation is obtained by adding these three values ​​together, representing the total energy consumed in performing a complete migration process.The net energy saving is calculated by subtracting the energy cost of the migration operation from the energy savings of the candidate data block. A positive net energy saving indicates that the energy saved after migration is greater than the energy consumed during the migration process, while a negative net energy saving indicates that the migration is not worthwhile. The average access latency time corresponding to the target storage medium type is retrieved from the storage medium performance parameter table. The access latency time represents the time required from initiating a read request to the return of data. Different storage media have access latency differences of orders of magnitude due to differences in physical characteristics. The net energy saving is divided by the access latency time to obtain the energy efficiency evaluation index. The energy efficiency evaluation index quantifies the energy saving effect under the unit access latency cost. The larger the value, the more energy saving is obtained under the acceptable increase in latency. The calculated energy efficiency evaluation index is written into the corresponding field of the candidate data block record in the database table.

[0057] Figure 2This is a schematic diagram comparing the energy consumption optimization effects of the IoT-based intelligent computing center in the embodiments of this application; the horizontal axis in the figure represents time (hours), from 0:00 to 24:00, the left vertical axis represents the energy consumption of the data center storage system (unit: kWh), ranging from 30-95 kWh, and the right vertical axis represents the real-time power generation of the photovoltaic power generation system (unit: kW), ranging from 0-60 kW. The figure contains three curves: The first curve, a black dashed line (connected by circular markers), represents the energy consumption change curve using the traditional fixed-time threshold determination method. This curve shows that the energy consumption of the traditional method remains high at 68-88 kWh during peak business hours (8:00-18:00); The second curve, a thick black solid line (connected by square markers), represents the energy consumption change curve using the method of intelligent prediction and renewable energy collaborative scheduling based on bidirectional long short-term memory networks proposed in this application. This curve is significantly lower than the curve of the traditional method throughout the day; The third curve, a thin dark gray solid line (connected by triangular markers, with a light gray area filled below the curve), represents the real-time dynamic change of the power generation of the data center rooftop photovoltaic power generation system. This curve shows a typical diurnal variation pattern, with power generation output during the period from 6:00 to 18:00, and the peak occurring around 12:00, reaching 52 kW. As can be clearly seen from the figure, the method of this application accurately predicts the probability of data block hot and cold state transitions and calculates energy consumption efficiency evaluation indicators. During periods of sufficient photovoltaic power generation (10:00-16:00, with photovoltaic power generation greater than 40kW), it intelligently schedules high-efficiency data blocks to perform migration tasks, effectively matching energy-intensive operations with the green energy supply window. Compared with traditional methods, it achieves significant energy consumption reduction effects at all times of the day. In particular, during peak business hours (8:00-18:00), the average energy consumption is reduced by about 18.2%, and during periods of sufficient photovoltaic power generation (10:00-16:00), the energy consumption reduction reaches 20.5%. This effectively verifies the superiority and practicality of the technical solution of this application in intelligently coordinating data migration scheduling with renewable energy supply status, while taking into account both energy savings and service quality assurance.

[0058] In one specific embodiment, step S4 includes:

[0059] Collect the real-time power generation of the photovoltaic power generation system in the data center and the remaining battery capacity, calculate the sum of the products of the two and the discharge efficiency, and obtain the renewable energy quota;

[0060] The renewable energy quota is compared with a preset high energy threshold, and the period when the renewable energy quota exceeds the preset high energy threshold is determined as the period when energy supply is sufficient.

[0061] During periods of sufficient energy supply, candidate data blocks are sorted in descending order according to energy consumption efficiency assessment indicators, and the ratio of renewable energy quota to average migration energy consumption of a single data block is calculated to obtain the upper limit of the number of data blocks that can be migrated.

[0062] Select data blocks from the sorted candidate data blocks that rank highly in energy efficiency evaluation indicators and whose number does not exceed the upper limit of the number of migrated data blocks, and construct a migration task sequence that includes data block identifier, source storage medium, target storage medium, and migration priority.

[0063] Specifically, power sensors deployed on the rooftop photovoltaic array of the data center collect the current output power of the photovoltaic power generation system in real time. This power value changes dynamically with the intensity of solar radiation and weather conditions. At the same time, the remaining battery capacity is read through the interface connected to the battery management system. The remaining battery capacity represents the total amount of usable electrical energy currently stored in the battery. Multiplying the remaining battery capacity by the discharge efficiency coefficient yields the actual releaseable power. The discharge efficiency coefficient reflects the energy conversion loss during the battery discharge process. The product represents the actual usable electrical energy after deducting the loss. The real-time power generation of the photovoltaic power generation system is added to the actual releaseable power of the battery to obtain the renewable energy quota. The renewable energy quota represents the total amount of green energy that the data center can use to perform data migration tasks at the current moment. This value comprehensively considers both on-demand power generation and energy storage release. The system reads a pre-set high-energy threshold parameter, which is a critical standard for determining the energy supply status. It compares the renewable energy quota value with the preset high-energy threshold. When the renewable energy quota exceeds the preset high-energy threshold, the current period is determined to be a period of sufficient energy supply. A period of sufficient energy supply indicates that green energy is abundant and suitable for performing energy-intensive tasks. During this period, data migration is performed to maximize the use of renewable energy and reduce the purchase of electricity from the grid. When the renewable energy quota is lower than the preset high-energy threshold, it is determined to be a period of energy constraint and not suitable for large-scale migration. After determining that the energy supply is sufficient, the energy efficiency evaluation index field of all candidate data blocks is read from the candidate data block list. The candidate data block list is then sorted. The sorting algorithm uses the energy efficiency evaluation index as the comparison key and sorts the data blocks in descending order of value. After sorting, the data block with the largest energy efficiency evaluation index is placed at the top of the list. These data blocks have the optimal energy saving to delay cost ratio. The average energy consumption of a single data block performing a complete migration operation is calculated from the historical migration task statistics as the average migration energy consumption of a single data block. This value is obtained by summing the actual energy consumption of all historical migration tasks and dividing by the total number of tasks. The current renewable energy quota value is divided by the average migration energy consumption of a single data block to obtain the upper limit of the number of data blocks that can be migrated. This upper limit represents the maximum number of data blocks that can be migrated under the current renewable energy constraints. The division result is rounded down to ensure that it does not exceed the energy budget.Starting from the first element in the sorted candidate data block list, data blocks are selected sequentially. During the selection process, the number of selected data blocks is accumulated. When the accumulated number reaches the upper limit of the number of migrated data blocks, the selection stops. The set of selected data blocks constitutes this migration batch. For each selected data block, its data block identifier field, current storage medium type field are queried from the original dataset as the source storage medium, and the target storage medium type to be migrated is used as the target storage medium. The migration priority is assigned according to the position of the data block in the sorted list, with higher priority values ​​for earlier positions. The four fields of data block identifier, source storage medium, target storage medium, and migration priority are combined into a migration task record. All migration task records of selected data blocks are arranged in priority order to form a migration task sequence. The migration task sequence is stored in a task queue data structure to await execution.

[0064] In one specific embodiment, step S5 includes:

[0065] Based on the data block identifiers and target storage media information in the migration task sequence, the data blocks are read, transferred, and written in descending order of migration priority to complete the media transfer of data blocks from the source storage media to the target storage media.

[0066] After the media transfer is completed, the actual access frequency of the migrated data blocks is continuously monitored within the preset monitoring period to obtain the actual access parameters of the migrated data blocks.

[0067] The difference between the actual access parameters and the predicted access parameters calculated based on the weighted access feature values ​​before migration is calculated to determine the relative deviation between the actual access parameters and the predicted access parameters.

[0068] When the relative deviation value exceeds the preset deviation threshold, the temporal feature sequence and actual access parameters of the corresponding data block are marked as correction samples, input into the bidirectional long short-term memory network for incremental learning and updating, and the preset decay coefficient in the time decay factor is adjusted.

[0069] Specifically, migration task records are retrieved sequentially from the task queue of the migration task sequence according to their migration priority field, from highest to lowest. For each migration task record, the data block identifier field, source storage medium field, and target storage medium field are extracted. Based on the data block identifier, the physical storage address of the data block on the source storage medium is queried in the storage system metadata table. A read command is sent to the block device driver of the source storage medium, containing the physical address and data length parameters. The source storage medium locates the data block storage location according to the command and reads the data content into the memory buffer to complete the read operation. After the read is completed, the data in the memory buffer is retrieved from the source storage medium via the data center internal network. The data is transmitted from the source storage medium to the target storage node. Network transmission ensures data integrity through the TCP protocol. After the data arrives at the target storage node, new physical storage space is allocated at the block device layer of the target storage medium. A write command is sent to the block device driver of the target storage medium. The write command contains the target physical address and data content. The target storage medium writes the data to the specified location according to the command to complete the write operation. After the write is completed, the storage system metadata table is updated to establish a mapping relationship between the data block identifier and the new physical address. At the same time, the metadata table is marked that the current storage medium type of the data block has been changed to the target storage medium type. Thus, the complete media transfer of the data block from the source storage medium to the target storage medium is completed. After the data block media transfer is completed, a continuous monitoring mechanism is started. A dedicated monitoring thread is created to track the access behavior of the migrated data blocks. The monitoring thread subscribes to access event notifications for the data block from the I / O monitoring interface of the storage system. Whenever the data block is accessed, the monitoring thread receives the event notification and records the timestamp of the access. The monitoring period is set to a preset duration starting from the time of migration completion. During the entire monitoring period, the monitoring thread continuously accumulates the total number of times the data block is accessed. At the end of the monitoring period, the total number of accesses is divided by the monitoring period duration to obtain the average access frequency of the data block after migration, i.e., the actual access parameter. The actual access parameter reflects the true access activity level of the data block after it is migrated to the target storage medium. The database is used to query the data block and obtain the future access frequency predicted by a bidirectional long short-term memory network based on the weighted access feature value before migration. This predicted access frequency is the access frequency estimated during the migration decision. The access frequency prediction error is obtained by subtracting the predicted access frequency value from the actual access frequency value. A positive prediction error indicates that the actual access frequency is higher than expected, while a negative value indicates that the actual access frequency is lower than expected. The absolute value of the prediction error is divided by the predicted access frequency to obtain the relative deviation value. The relative deviation value is a dimensionless relative error index that reflects the accuracy of the prediction.The calculated relative deviation value is compared with a preset deviation threshold, which is the standard for determining whether the prediction error is too large. When the relative deviation value exceeds the preset deviation threshold, it is determined that the prediction of the data block has a significant error and the prediction model needs to be corrected. The weighted access feature value sequence of the data block used for prediction within the continuous time window before migration is retrieved from the database as the time-series feature sequence. At the same time, the actual access frequency of the data block within the monitoring period after migration is extracted as the true label. The time-series feature sequence is combined as input feature and the true label as expected output to form a pair of training samples. The training sample is marked as a correction sample and stored in the correction sample set. When the correction sample set accumulates to a preset number, the model update process is triggered. The correction sample set is input into the bidirectional long short-term memory network in batches for incremental learning. During the incremental learning process, the network reads the time-series feature sequence input of the correction sample and calculates the prediction output through the forward and backward long short-term memory layers. The prediction output is then compared with the true label. The loss function is calculated, and its gradient with respect to the network weight parameters is determined using the backpropagation algorithm. The network weight parameters are then adjusted along the gradient descent direction. The learning rate for incremental learning is set to a small multiple of the initial training learning rate to avoid forgetting learned knowledge. After multiple iterations, the network weight parameters are updated. Simultaneously, the ratio of actual transfer energy consumption to predicted transfer energy consumption for all transferred tasks executed in the past time period is calculated. The arithmetic mean of all ratios is calculated to obtain the energy consumption calibration coefficient. The energy consumption calibration coefficient reflects the systematic bias of the energy consumption prediction model. The energy consumption calibration coefficient is compared with a preset benchmark value. When the absolute value of the difference between the two exceeds a preset calibration threshold, it is determined that there is a significant bias in the energy consumption model. The preset decay coefficient in the time decay factor calculation formula is multiplied by the energy consumption calibration coefficient to obtain the adjusted decay coefficient. The adjusted decay coefficient replaces the original preset decay coefficient to update the calculation parameters of the time decay factor. The updated decay coefficient takes effect in the weighted access feature value calculation of subsequent data blocks.

[0070] In one specific embodiment, when the relative deviation value exceeds a preset deviation threshold, the temporal feature sequence and actual access parameters of the corresponding data block are marked as correction samples, input into the bidirectional long short-term memory network for incremental learning and updating, and the preset decay coefficient in the time decay factor is adjusted, including:

[0071] Extract the weighted access feature value sequence and actual access frequency within the continuous time window corresponding to the data block whose relative deviation value exceeds the preset deviation threshold, and combine them into a calibration sample set containing the input feature sequence and the true label;

[0072] The calibration sample set is input into the bidirectional long short-term memory network, the learning rate is set to one-tenth of the initial training learning rate, and the network weight parameters are updated iteratively by gradient descent.

[0073] After incremental learning is completed, the ratio of actual migration energy consumption to predicted migration energy consumption within a preset time period is calculated to obtain the energy consumption calibration coefficient.

[0074] When the deviation between the energy consumption calibration coefficient and the preset benchmark value exceeds the preset calibration threshold, the preset attenuation coefficient is adjusted proportionally according to the energy consumption calibration coefficient, and the calculation parameters of the time attenuation factor are updated.

[0075] Specifically, all data block records whose relative deviation value exceeds a preset deviation threshold are selected from the monitoring record table of the database. For each selected record, the corresponding data block identifier is extracted. Based on the data block identifier, the weighted access feature value sequence used for prediction within a continuous time window before migration is retrieved from the historical data table. This sequence contains weighted access feature values ​​at multiple time points arranged in chronological order. At the same time, the actual access frequency value recorded within the monitoring period after migration of the data block is extracted from the monitoring record table. The weighted access feature value sequence is used as the input feature sequence, and the actual access frequency value is used as the ground truth label. The two are paired to form a complete training sample. In the training sample, the input feature sequence is used for the network to learn access patterns, and the ground truth label is used to guide the network to adjust the prediction direction. The training samples corresponding to all selected data blocks are summarized and stored in the calibration sample set data structure. The calibration sample set contains multiple such input-output paired samples. The training samples from the calibration sample set are input into a bidirectional long short-term memory network in batches. Each batch contains several training samples. The network reads the input feature sequence of each sample in the batch and processes the weighted access feature values ​​step by step from the start position to the end position through a forward long short-term memory layer. At the same time, it processes the sequence in reverse from the end position to the start position through a backward long short-term memory layer. The hidden states in both directions are concatenated at each time step and then passed through a fully connected layer and an activation function layer to output the predicted access frequency. The mean squared error is calculated by comparing the predicted access frequency output by the network with the true label in the training samples, i.e., the actual access frequency, as the loss function. The smaller the loss function value, the more accurate the prediction. The backpropagation algorithm is then used. The gradient is obtained by calculating the partial derivative of the loss function with respect to each weight parameter in the network. The gradient indicates the direction and magnitude of the weight parameter adjustment. Each weight parameter is moved by a certain step size in the negative direction of its gradient to complete the parameter update. The step size is controlled by the learning rate, which is set to one-tenth of the learning rate at the beginning of training. A smaller learning rate results in a smaller parameter adjustment and avoids destructive overwriting of the network's learned historical access pattern knowledge. The process of forward propagation to calculate the loss and backpropagation to update the parameters is repeated for all batches of training samples to complete one round of iteration. After multiple rounds of iteration, the network weight parameters converge to a new optimal value. At this point, the network has a more accurate predictive ability for access patterns that include samples with prediction errors.After incremental learning is completed, all completed data migration task records within a preset time period are queried from the migration task execution log table. The preset time period is the time window for statistical energy consumption deviation. For each migration task record, the actual migration energy consumption field and the predicted migration energy consumption field are extracted. The actual migration energy consumption is the total energy consumption measured by the power consumption sensor during the execution of the task. The predicted migration energy consumption is the energy consumption estimate calculated before the task execution based on the data block size and storage medium parameters. The actual migration energy consumption value is divided by the predicted migration energy consumption value to obtain the energy consumption ratio of the task. The arithmetic mean of the energy consumption ratios of all migration tasks within the preset time period is used to obtain the energy consumption calibration coefficient. The energy consumption calibration coefficient reflects the degree of systematic deviation of the energy consumption prediction model. A ratio greater than one indicates that the actual energy consumption is generally higher than the prediction and the migration cost is underestimated. A ratio less than one indicates that the actual energy consumption is generally lower than the prediction and the migration cost is overestimated. The system reads the preset baseline value parameter, which is set to an ideal state where the actual energy consumption and the predicted energy consumption are completely consistent. It calculates the deviation by subtracting the preset baseline value from the energy consumption calibration coefficient. The absolute value of the deviation is compared with the preset calibration threshold, which is the critical standard for determining whether the energy consumption model needs calibration. When the absolute value of the deviation exceeds the preset calibration threshold, it is determined that there is a significant systematic error in the energy consumption prediction and the model parameters need to be adjusted. The system reads the preset decay coefficient used in the current time decay factor calculation formula and multiplies the preset decay coefficient by the energy consumption calibration coefficient to obtain the adjusted decay coefficient. The adjusted decay coefficient reflects the impact of the energy consumption prediction deviation on the data activity judgment. The system replaces the original preset decay coefficient with the adjusted decay coefficient in the updated configuration parameter table. The updated decay coefficient takes effect in the subsequent calculation of the weighted access feature value of the data block, affecting the value of the time decay factor and thus changing the judgment of the data hot and cold status.

[0076] The above describes the energy consumption optimization method for an IoT-based intelligent computing center in the embodiments of this application. The following describes the energy consumption optimization system for an IoT-based intelligent computing center in the embodiments of this application. Please refer to [link / reference]. Figure 3 One embodiment of the IoT-based intelligent computing center energy consumption optimization system in this application includes:

[0077] The generation module is used to collect energy consumption characteristic parameters of different types of storage media in the data center and data block access behavior parameters through the Internet of Things sensor network, and generate a raw dataset containing time information, data block identifiers and energy consumption characteristics.

[0078] The weighting module is used to perform weighted calculations on the access behavior parameters and time decay factors in the original dataset to obtain access feature values, and to calculate the cold and hot state transition probability of the data block based on the access feature values;

[0079] The calculation module is used to filter candidate data blocks whose cold and hot state transition probabilities meet the set conditions, calculate the energy consumption difference between the candidate data blocks and the target storage medium and the energy consumption cost of the migration operation, and obtain the energy consumption benefit evaluation index.

[0080] The matching module is used to match and judge the energy consumption efficiency assessment indicators with the renewable energy supply status, and select high-efficiency data blocks to construct a migration task sequence during periods of sufficient energy supply.

[0081] The acquisition module is used to perform media transfer operations on data blocks in the migration task sequence, and to acquire the deviation information between the actual access parameters and the predicted access parameters of the data blocks after migration for dynamic adjustment of access feature value calculation parameters.

[0082] above Figure 3 The energy consumption optimization system for an IoT-based intelligent computing center in this embodiment of the invention will be described in detail from the perspective of modular functional entities. The energy consumption optimization device for an IoT-based intelligent computing center in this embodiment of the invention will be described in detail from the perspective of hardware processing.

[0083] Reference Figure 4 This invention also provides an IoT-based intelligent computing center energy consumption optimization device, which can be a server, and its internal structure can be as follows: Figure 4 As shown, this IoT-based intelligent computing center energy optimization device includes a processor, memory, display screen, input device, network interface, and database connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The database stores the data corresponding to this embodiment. The network interface communicates with external terminals via a network connection. When the computer program is executed by the processor, it implements the aforementioned method.

[0084] Those skilled in the art will understand that Figure 4 The structure shown is merely a block diagram of a portion of the structure related to the present invention and does not constitute a limitation on the energy consumption optimization device for the IoT-based intelligent computing center on which the present invention is applied.

[0085] The present invention also provides a computer-readable storage medium, which can be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium. The computer-readable storage medium stores instructions that, when executed on a computer, cause the computer to perform the steps of the IoT-based intelligent computing center energy consumption optimization method.

[0086] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0087] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause an IoT-based intelligent computing center energy optimization device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0088] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An Internet of Things-based intelligent computing center energy consumption optimization method, characterized in that, The method comprises the following steps: S1, collecting energy consumption characteristic parameters of different types of storage media and access behavior parameters of data blocks in a data center through an Internet of Things sensor network, and generating an original data set containing time information, data block identification and energy consumption characteristics; S2, performing weighted operation on the access behavior parameters in the original data set and a time decay factor to obtain access characteristic values, and calculating the cold-hot state conversion probability of the data blocks based on the access characteristic values; S3, comparing the cold-hot state conversion probability with a preset probability threshold, screening data blocks with a cold-hot state conversion probability greater than the preset probability threshold as candidate data blocks; extracting the current storage medium unit power consumption value and the target storage medium unit power consumption value of the candidate data blocks from the original data set, calculating the product of the power consumption difference value and the predicted cold state retention time of the data blocks, and obtaining the energy consumption saving amount of the candidate data blocks; calculating the sum of the read energy consumption, network transmission energy consumption and write energy consumption of the candidate data blocks to obtain the migration operation energy consumption cost of the candidate data blocks from the current storage medium to the target storage medium; performing difference operation on the energy consumption saving amount and the migration operation energy consumption cost, and then dividing by the data block access delay time to obtain the energy consumption benefit evaluation index of the candidate data blocks; S4, matching and judging the energy consumption benefit evaluation index and the renewable energy supply state, selecting high-benefit data blocks during the period of sufficient energy supply to construct a migration task sequence; S5, according to the data block identification and target storage medium information in the migration task sequence, performing read, transmission and write operations on the data blocks in sequence according to the migration priority from high to low, and completing the medium transfer of the data blocks from the source storage medium to the target storage medium; After the medium transfer is completed, the actual access frequency of the migrated data blocks is continuously monitored within a preset monitoring period to obtain the actual access parameters of the data blocks after migration; performing difference operation on the actual access parameters and the predicted access parameters calculated based on the weighted access characteristic values before migration to calculate the relative deviation value of the actual access parameters and the predicted access parameters; when the relative deviation value exceeds a preset deviation threshold, the time sequence feature sequence and the actual access parameters of the corresponding data blocks are marked as correction samples, which are input into a bidirectional long short-term memory network for incremental learning and updating to adjust the preset decay coefficient in the time decay factor. 2.The IoT-based intelligent computing center energy consumption optimization method of claim 1, wherein, S1 comprises: Deploying a power consumption monitoring module in the storage cabinet of the data center, collecting the unit power consumption values of the disk array, solid state disk and blue light storage medium to obtain a storage medium energy consumption parameter set; recording the access frequency, access time interval and data block size of each data block at the time stamp moment through a data access behavior monitoring sensor to obtain an access behavior parameter set; associating and matching the storage medium energy consumption parameter set and the access behavior parameter set according to the data block identification to generate an associated data record containing the time stamp, data block identification, access frequency, current storage medium type and real-time power consumption value; normalizing the access frequency in the associated data record, dividing the access frequency by the maximum access frequency value to map to a preset numerical interval to obtain an original data set. 3.The IoT-based intelligent computing center energy consumption optimization method of claim 1, wherein, S2 comprises: extracting the access time interval of each data block from the original data set, performing negative exponential operation on the access time interval and the preset decay coefficient to obtain the natural constant power, and obtaining the time decay factor of each data block; multiplying the normalized access frequency in the original data set and the corresponding time decay factor to obtain the weighted access characteristic value reflecting the actual activity degree of the data block; extracting the weighted access characteristic value in the continuous time window in the original data set to construct the time sequence characteristic sequence, inputting the time sequence characteristic sequence into the bidirectional long short-term memory network for forward and backward time sequence feature extraction to obtain the implicit state vector of the data block access trend; the implicit state vector is mapped through the activation function layer to output the cold and hot state conversion probability of the data block from hot state to cold state in the future time window. 4.The IoT-based intelligent computing center energy consumption optimization method of claim 1, wherein, S4 comprises: Collecting the real-time power generation of the data center photovoltaic power generation system and the remaining capacity of the storage battery, calculating the sum of the product of the two and the discharge efficiency to obtain the renewable energy quota; comparing the renewable energy quota with the preset high energy threshold to determine the period when the renewable energy quota exceeds the preset high energy threshold as the energy supply sufficient period; in the energy supply sufficient period, the candidate data blocks are arranged in descending order according to the energy consumption benefit evaluation index, and the ratio of the renewable energy quota to the average migration energy consumption of a single data block is calculated to obtain the upper limit of the number of migratable data blocks; Selecting data blocks with high energy consumption benefit evaluation index from the sorted candidate data blocks and not more than the upper limit of the number of migratable data blocks to construct a migration task sequence containing data block identification, source storage medium, target storage medium and migration priority. 5.The IoT-based intelligent computing center energy consumption optimization method of claim 1, wherein, When the relative deviation value exceeds the preset deviation threshold, the time sequence characteristic sequence and the actual access parameter of the corresponding data block are marked as correction samples, input into the bidirectional long short-term memory network for incremental learning update, and the preset decay coefficient in the time decay factor is adjusted, including: Extracting the weighted access characteristic value sequence and the actual access frequency in the continuous time window corresponding to the data block whose relative deviation value exceeds the preset deviation threshold to form a correction sample set containing input feature sequence and true label; Input the correction sample set into the bidirectional long short-term memory network, set the learning rate to one tenth of the initial training learning rate, and iteratively update the network weight parameters by gradient descent; After the incremental learning is completed, the ratio of the actual migration energy consumption to the predicted migration energy consumption in the preset time period is calculated to obtain the energy consumption calibration coefficient; When the deviation of the energy consumption calibration coefficient from the preset reference value exceeds the preset calibration threshold, the preset decay coefficient is adjusted in proportion according to the energy consumption calibration coefficient, and the calculation parameters of the time decay factor are updated.

6. An Internet of Things based intelligent computing center energy consumption optimization system, characterized in that, The system for implementing the method for optimizing the energy consumption of the intelligent computing center based on the Internet of Things comprises: A generation module is configured to collect energy consumption characteristic parameters of different types of storage media and access behavior parameters of data blocks in a data center through an Internet of Things sensor network, and generate an original data set containing time information, data block identification and energy consumption characteristics; The weighting module is configured to perform weighting operation on the access behavior parameters in the original data set and the time decay factor to obtain access characteristic values, and calculate the cold-hot state transition probability of the data block based on the access characteristic values; The computing module is configured to filter candidate data blocks whose cold-hot state transition probability meets a set condition, calculate a comparison value of the energy consumption difference between the current storage medium and the target storage medium and the energy consumption cost of the migration operation, and obtain an energy consumption benefit evaluation index; The matching module is configured to match the energy consumption benefit evaluation index with the renewable energy supply state, and select high-benefit data blocks to construct a migration task sequence during an energy supply sufficient period; The acquisition module is configured to perform medium migration operation on the data blocks in the migration task sequence, and acquire deviation information of the actual access parameters and the predicted access parameters of the migrated data blocks for dynamic adjustment of the access characteristic value calculation parameters.

7. An Internet of Things based intelligent computing center energy consumption optimization device, characterized in that, The computer program is run on the processor to implement the method for optimizing energy consumption of an intelligent computing center based on Internet of Things according to any one of claims 1 to 5.

8. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is run on the processor to implement the method for optimizing energy consumption of an intelligent computing center based on Internet of Things according to any one of claims 1 to 5.

Images