Building energy consumption prediction method based on artificial intelligence

Abstract

The invention relates to the technical field of building energy consumption prediction, in particular to a building energy consumption prediction method based on artificial intelligence, and the method comprises the following steps: removing repeated entries, removing abnormal records and reconstructing a time dimension when data cleaning and time division are carried out based on a floor use condition and a sensor monitoring numerical value. According to the method, through fine data cleaning and time period reconstruction based on floor use conditions and sensor values, abnormal records are removed, data measurement units are unified, and accuracy and consistency of energy consumption data are ensured; equipment operation records and maintenance items in the maintenance log are associated, the coupling degree of operation data and actual maintenance operation is enhanced, and the integrity and reliability of data information are improved; through text word segmentation and event labeling, an accurate analysis date range is delimited, noise information is rejected, building management and control nodes are linked, and the recognition efficiency and accuracy of operation and maintenance information are enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of building energy consumption prediction, and in particular to a building energy consumption prediction method based on artificial intelligence. Background Art

[0002] The technical field of building energy consumption prediction mainly focuses on the electricity consumption, heat consumption and cooling supply conditions during the operation of buildings. By using data analysis and energy consumption management means, and with the help of building automation, sensor networks and data acquisition devices, various types of data including building structure information, environmental climate factors, personnel activity patterns, and equipment operation parameters are obtained to analyze the energy consumption trends and characteristics of buildings.

[0003] In the prior art, there is a lack of effective integration of operation and maintenance logs and equipment maintenance matters. Most of the operation and maintenance information is unstructured text, and the data is not accurately associated with actual operations, making the corresponding relationship between building operation and maintenance events and actual energy consumption trends unclear. In addition, the prior art fails to conduct differential segmented analysis for different load periods and accurately locate high-energy-consuming nodes, and does not monitor and quantitatively analyze the current fluctuations and power change trends of equipment, resulting in difficulty in promptly responding to load fluctuations and dynamically adjusting control strategies, affecting the timeliness and accuracy of prediction results. For example, during equipment maintenance, failure to adjust control strategies in a timely manner will cause energy waste, reduce energy utilization efficiency, and increase operating costs. Summary of the Invention

[0004] The purpose of the present invention is to solve the deficiencies existing in the prior art, and a building energy consumption prediction method based on artificial intelligence is proposed.

[0005] To achieve the above purpose, the present invention adopts the following technical solutions. A building energy consumption prediction method based on artificial intelligence includes the following steps:

[0006] Based on the floor usage and sensor monitoring values, when performing data cleaning and time period division, duplicate entries are removed, abnormal records are eliminated, and the time dimension is reconstructed. When screening temperature, humidity, and power information, the measurement units are unified and the timestamps are calibrated. The operation records of the heat exchanger and the maintenance matters of the exhaust equipment in the maintenance log are associated to obtain a multi-source splicing sequence.

[0007] Based on the multi-source splicing sequence, text segmentation and operation event annotation are performed, the date range is delimited, noise words are filtered, the building management and control nodes in the knowledge graph are linked, the floor numbers and operation instruction types are identified, and the operation instructions and sensor note records in the maintenance log are integrated to obtain an operation and maintenance integration information set.

[0008] Based on the operation and maintenance integration information set, perform time series splitting and difference calculation, detect missing intervals and fill in breakpoint data, cross-compare temperature differences when matching floor usage and external meteorological indicators, summarize the usage rate classification records in the operation file, screen the load change intervals, and obtain the coupling feature matrix;

[0009] Based on the coupling feature matrix, perform segmented parsing and time period association matching, split high-frequency segments and merge discontinuous time periods, record the change trends when monitoring the indoor air circulation rate and the floor lighting start signal, verify the filter blockage records in the maintenance log, identify the energy-consuming nodes and quantify the fluctuation intensity, and obtain the energy consumption deviation parameters.

[0010] Preferably, it further includes: based on the energy consumption deviation parameters, perform real-time monitoring data comparison and difference push, extract the floor equipment current fluctuation and record the peak value, compare the changes in floor usage when adjusting the input signal of the building control terminal and correcting the energy distribution ratio, merge the operation period information in the maintenance log, output the control and scheduling list and then integrate the status weights to obtain the adaptive calibration matrix;

[0011] Based on the adaptive calibration matrix, accumulate the sampling differences of floor temperature and humidity, fuse the floor usage and meteorological data and statistically analyze the distribution characteristics, synchronously record the wind speed elements, analyze the energy consumption trends of each node in the control and scheduling list and output the predicted time series fluctuation curve to obtain the energy consumption prediction sequence.

[0012] Preferably, the steps for obtaining the multi-source splicing sequence are as follows:

[0013] Based on the floor usage and sensor monitoring values, perform record verification and duplicate entry screening to eliminate the same records and remove damaged data, reorganize the time series and label the monitoring periods, exclude abnormal dates and synchronously align the floor numbers to obtain the time period corrected data;

[0014] Based on the time period corrected data, perform temperature, humidity and power field screening to unify the measurement units and align the sampling frequencies, eliminate invalid measuring points and label the floor positions, associate the operation periods of the heat exchangers in the maintenance log and compare with the exhaust equipment maintenance records to obtain the screened temperature, humidity and power group;

[0015] Based on the screened temperature, humidity and power group, perform missing period verification and merge the operation records, reconstruct the time series index and allocate the floor coding fields, stack the adjacent period states and generate the overall building mapping, check the duplicate sampling identifiers and summarize the collaborative operation information of the heat exchangers and exhaust equipment to obtain the multi-source splicing sequence.

[0016] Preferably, the steps for obtaining the operation and maintenance integration information set are as follows:

[0017] Based on the multi-source splicing sequence, floor numbers and date tags are extracted during text parsing and event splitting, irrelevant entries are screened out, and the operation status description of the maintenance log is matched. The building control node information is associated, and the source of the operation instruction is identified to obtain an event annotation group;

[0018] Based on the event annotation group, floor field comparison is carried out to map and maintain the inspection records in the log. Duplicate instructions are removed, and records of the same type in the same time period are merged. Noise words are identified, and a floor operation entry table is established. The date tags are proofread, and the operation instruction types are associated to obtain an operation and maintenance dictionary set;

[0019] Based on the operation and maintenance dictionary set, the instruction sources are classified, and the sensor note fields are matched. The floor usage status and key operation descriptions are split, the building control nodes are linked, and the operation event contents are merged. The text information of all floors is summarized, and the time series tags are sorted to obtain an operation and maintenance integration information set.

[0020] Preferably, the steps for obtaining the coupling feature matrix are as follows:

[0021] Based on the operation and maintenance integration information set, time periods are split, and the continuity of the dates is checked. Data breakpoints are filled, and the missing intervals are recorded. The floor usage conditions are read, and the external meteorological values are compared. The floor occupancy rate is analyzed, and the temperature difference is calculated. After marking the cross fields, time series difference data is obtained;

[0022] Based on the time series difference data, the missing intervals are located, and the classified records of the floor usage rate are matched. The power peak and valley areas are screened, and the change range of the meteorological values is analyzed. Combined with the temperature difference and time period marking, the integrity of the records is checked, and an interval index group is generated to obtain a load change interval group;

[0023] Based on the load change interval group, coupling analysis is carried out, and the floor usage records are matched. The load peak value is retrieved, and the correlation between the temperature, humidity, and power changes is located. The time periods are extracted, and the occupancy rate and meteorological data are merged. After marking the load fluctuation characteristics, a coupling feature matrix is obtained.

[0024] Preferably, the steps for obtaining the energy consumption deviation parameter are as follows:

[0025] Based on the coupling feature matrix, segmented parsing is carried out, and high-frequency paragraphs are retrieved. Time windows are divided, and the indoor air flow rate and the floor lighting start signal are aligned. The change trend and the record of filter blockage in the maintenance log are recorded. After checking the floor numbers and energy consumption records, segmented correlation information is obtained;

[0026] Based on the segmented correlation information, the floor lighting status and the indoor temperature gradient are monitored, and the lighting on-off records and the temperature and humidity increments are compared. Abnormal sections are extracted, and the filter blockage periods are checked. The high-frequency operation points of the exhaust equipment are retrieved, and the energy consumption peak value is quantified to obtain a high-energy consumption node group;

[0027] Based on the high - energy - consuming node group, read the peak period of the lighting equipment and match the synchronous information of the heat exchanger, superimpose the floor temperature and humidity values and air velocity and compare the power fluctuation trajectory, check the floor number matching and divide the fluctuation intensity level, and obtain the energy consumption deviation parameter after merging the abnormal peak segments.

[0028] Preferably, the steps for obtaining the adaptive calibration matrix are as follows:

[0029] Based on the energy consumption deviation parameter, extract the peak current of the floor equipment and retrieve the power fluctuation period, compare with the indoor temperature and humidity and the operation information of the exhaust equipment, count the deviation data and mark the difference level, aggregate the changes in floor usage and record the corresponding time period, and obtain the monitoring and comparison record;

[0030] Based on the monitoring and comparison record, read the input signal of the building control terminal and screen the energy distribution instruction, retrieve the changes in floor usage and check the scheduling period of the exhaust equipment, superimpose the operation instruction records in the maintenance log and analyze the current peak and valley, and obtain the scheduling list group;

[0031] Based on the scheduling list group, merge the operation instruction and the status weight and analyze the changes in floor usage, verify the energy rationing ratio and record the calibration time on the sensor side, check the response of the building control terminal and summarize the floor temperature and humidity deviation and current fluctuation information, and obtain the adaptive calibration matrix.

[0032] Preferably, the steps for obtaining the energy consumption prediction sequence are as follows:

[0033] Based on the adaptive calibration matrix, accumulate the sampling difference of the floor temperature and humidity and verify the date consistency, select the meteorological values and wind speed elements and count the corresponding distribution, mark the switching periods of the floor usage and the indoor operation state, merge the maintenance log and compare the operation information of the exhaust equipment, and obtain the iterative difference record;

[0034] Based on the iterative difference record, analyze the coupling relationship between the floor usage and the meteorological values and estimate the temperature and humidity change rate, retrieve the current peak and valley segments and compare the energy rationing ratio, count the distribution of the operation periods of each floor and mark the abnormal fluctuation information, and obtain the fluctuation curve group;

[0035] Based on the fluctuation curve group, read the floor occupancy rate and meteorological data and proofread the temperature and humidity difference, verify the exhaust equipment statistics and summarize the node energy consumption changes in the management and control scheduling list and then analyze the floor power trend, screen the key time periods and draw the time - series trend chart, and obtain the energy consumption prediction sequence.

[0036] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0037] Through fine data cleaning and time period reconstruction based on floor usage and sensor values, the present invention removes abnormal records and unifies data measurement units, ensuring the accuracy and consistency of energy consumption data; associates and maintains the equipment operation records and maintenance matters in the log, strengthens the coupling degree between operation data and actual maintenance operations, and improves the integrity and reliability of data information; through text segmentation and event annotation, delimits the accurate analysis date range, eliminates noise information and links building control nodes, enhancing the recognition efficiency and accuracy of operation and maintenance information; through difference calculation and filling of breakpoint data, cross-compares indoor and outdoor temperature, humidity and meteorological changes, and clarifies the dynamic response characteristics between floor operation load and external environment; through segmental analysis of high-frequency energy consumption sections, monitors the change trends of indoor air velocity and equipment opening and closing signals, quickly locates key energy-consuming nodes and quantifies the fluctuation intensity, optimizing the real-time monitoring level of high-energy-consuming nodes; through monitoring of current fluctuation peaks and adjustment of energy distribution ratios, integrates multi-dimensional information of maintenance logs and equipment operation states, dynamically outputs a control and scheduling list, enhancing the real-time and adaptive capabilities of energy consumption control; through integrating floor usage, meteorological data and wind speed elements, accumulatively analyzes the distribution characteristics of temperature and humidity differences, analyzes the energy consumption trends of each node and outputs a predicted fluctuation curve, realizing the efficient prediction and tracking of energy consumption trends and comprehensively improving the building energy consumption prediction effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 It is a schematic diagram of the steps of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0039] In order to make the objectives, technical solutions and advantages of the present invention clearer and more understandable, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0040] Please refer to Figure 1 , the present invention provides a technical solution, an artificial intelligence-based building energy consumption prediction method, including the following steps:

[0041] When performing data cleaning and time period division based on floor usage and sensor monitoring values, duplicate entries are removed, abnormal records are eliminated, and the time dimension is reconstructed. When screening temperature, humidity and power information, the measurement units are unified and the timestamps are calibrated. The operation records of the heat exchanger and the maintenance matters of the exhaust equipment in the maintenance log are associated to obtain a multi-source splicing sequence;

[0042] Based on the multi-source splicing sequence, text segmentation and operation event annotation are performed, the date range is delimited and noise words are filtered, the building control nodes in the knowledge graph are linked, the floor numbers and operation instruction types are identified, and the operation instructions and sensor note records in the maintenance log are integrated to obtain an operation and maintenance integration information set;

[0043] Based on the operation and maintenance integration information set, conduct time series splitting and difference calculation, detect missing intervals and fill in breakpoint data, cross-compare temperature differences when matching floor usage and external meteorological indicators, summarize the usage rate classification records in the operation archives, screen the load change intervals, and obtain the coupling feature matrix;

[0044] Based on the coupling feature matrix, conduct segmented analysis and time period correlation matching, split high-frequency sections and merge discontinuous time periods, record the change trends when monitoring the indoor air circulation rate and floor lighting start signals, verify the filter clogging records in the maintenance log, identify energy-consuming nodes and quantify the fluctuation intensity, and obtain the energy consumption deviation parameters.

[0045] Based on the energy consumption deviation parameters, conduct real-time monitoring data comparison and difference push, extract the current fluctuations of floor equipment and record the peak values, compare the changes in floor usage when adjusting the input signals of the building control terminal and correcting the energy distribution ratio, merge the operation period information in the maintenance log, and integrate the status weights after outputting the control and scheduling list to obtain the adaptive calibration matrix;

[0046] Based on the adaptive calibration matrix, accumulate the sampling differences of floor temperature and humidity, fuse the floor usage and meteorological data and statistically analyze the distribution characteristics, synchronously record the wind speed elements, analyze the energy consumption trends of each node in the control and scheduling list and output the predicted time series fluctuation curve to obtain the energy consumption prediction sequence.

[0047] The steps for obtaining the multi-source splicing sequence are as follows:

[0048] Based on the floor usage and sensor monitoring values, conduct record verification and duplicate entry investigation, eliminate the same records and remove damaged data, reorganize the time series and mark the monitoring time periods, exclude abnormal dates and synchronously align the floor numbers to obtain the time period corrected data;

[0049] Based on the time period corrected data, conduct temperature, humidity and power field screening, unify the measurement units and align the sampling frequencies, eliminate invalid measuring points and mark the floor positions, associate the operation time periods of the heat exchanger in the maintenance log and compare with the exhaust equipment maintenance records to obtain the screened temperature, humidity and power group;

[0050] Based on the screened temperature, humidity and power group, conduct missing time period verification and merge the operation records, reconstruct the time series index and allocate the floor coding fields, stack the adjacent time period states and generate the overall building mapping, check the duplicate sampling identifiers and summarize the collaborative operation information of the heat exchanger and the exhaust equipment to obtain the multi-source splicing sequence.

[0051] Specifically, according to the floor usage and sensor monitoring values ​​obtained previously, the corresponding sequences are first extracted from the basic information such as temperature, humidity and power, and rearranged at fixed time intervals, and then the timestamp and floor number of each record are compared. If the timestamp is repeated or missing, it is interpolated and supplemented according to the time difference of adjacent data. The temperature value of each record is compared with the range of 0℃ to 90℃, the humidity value is compared with the range of 0% to 100%, and the power value is compared with the range of 0W to 5000W. All those exceeding the range are marked as suspicious entries and manually reviewed. The entries that are confirmed to have extreme deviations after review are completely eliminated. Then, a three-layer fully connected neural network is used to determine whether there are still abnormal data and generate the final valid record set. The input vector of the neural network includes three types of values ​​​​of temperature, humidity and power and the corresponding time interval information. The hidden layer contains two layers of 100 neurons each and uses ReLU as the activation function. The output layer uses the Sigmoid function to determine whether the record is abnormal. During the training process, normal records and abnormal records that have been confirmed in the early stage are selected as training sets. To measure the model error, a learning rate of 0.001 is used and 200 rounds are iterated each time the weight is updated. After training, new records are sent to the network for inference. If the output value is greater than 0.8, the record is judged to be abnormal, otherwise it is judged to be normal. When all the data have been judged and processed, a continuous time period baseline is generated according to the chronological order of the records and the floor number is matched with the valid observation value to finally obtain the time period correction data.

[0052] According to the period-corrected data obtained earlier, the temperature, humidity and power fields are read and unified in Celsius, relative humidity percentage and watt as the units of measurement. A comparison table is used to convert the possible Fahrenheit temperature or kilowatt values ​​into the corresponding standard format. In order to identify whether there are short-term drastic fluctuations in the period-corrected data, the temperature of each record is compared with the range of 0℃ to 90℃, the humidity is compared with the range of 0% to 100%, and the power is compared with the range of 0W to 5000W. If an entry exceeds the limit multiple times in a short period of time, it is marked with a special mark, and a set of deep convolutional neural networks is used for multi-dimensional screening. The network contains four convolutional layers and two fully connected layers. The input part is a two-dimensional tensor composed of temperature, humidity, power and time series. The convolution kernel size is set to 3×3. The activation function uses LeakyReLU and batch normalization is added after each layer to maintain numerical stability. During training, the labeled normal segments and abnormal segments are extracted as data sets, and the cross entropy loss function is used. Update the weights with a learning rate set to 0.0005 and iterate for 300 rounds. After training is completed, perform feature extraction and classification on the new input. If the output probability exceeds 0.7, it is considered that the field combination for this period requires manual review again. If the probability is lower than this value, it is determined as acceptable data. For the determined acceptable data, compare it with the time intervals of the heat exchanger operation and the exhaust equipment maintenance in the maintenance log, retain the records where both are within the valid range, and summarize them into a set of three fields: temperature, humidity, and power, finally obtaining the filtered temperature, humidity, and power group.

[0053] According to the filtered temperature, humidity, and power group obtained previously, recheck the possible missing periods in its records. If it is found that the temperature, humidity, or power is completely missing in some time segments, fill them in by interpolation, and verify the rationality of the filled data through a time series prediction model based on gated recurrent units. The input of this model includes the temperature sequence, humidity sequence, power sequence, and the corresponding time index. The hidden layer has two gated recurrent unit layers, and each gated unit contains several update gates and reset gates to perform operations on the sequence in turn. During training, split continuous segments from the existing complete records as the training set, and use the mean absolute error to update the model weights, set the learning rate to 0.001 and perform 150 rounds of iteration. Subsequently, input the sequence after interpolation into this model for forward and backward prediction comparison. If the deviation between the predicted value and the filled value exceeds the set threshold of 0.15, it is regarded as an unreasonable missing filling and handed over to manual item-by-item inspection. For the periods with confirmed reasonable filling, superimpose the floor coding field with the adjacent period status and perform sequence splicing in an overall mapping. Finally, check all duplicate sampling identifiers and summarize the collaborative operation information of the heat exchanger and the exhaust equipment, ultimately obtaining the multi-source splicing sequence.

[0054] The steps to obtain the operation and maintenance integration information set are as follows:

[0055] Based on the multi-source splicing sequence, extract the floor number and date label during text parsing and event splitting, screen out irrelevant entries, match the operation status description in the maintenance log, associate the building control node information, and identify the source of the operation instruction to obtain the event annotation group;

[0056] Based on the event annotation group, perform floor field comparison and map the maintenance log records for maintenance, eliminate duplicate instructions, merge records of the same type in the same period, identify noise words, and establish a floor operation entry table, proofread the date label, and associate the operation instruction type to obtain the operation and maintenance dictionary set;

[0057] Based on the operation and maintenance dictionary set, classify the source of the instruction, match the sensor note field, split the floor usage status and key operation description, link the building control node, and merge the operation event content, summarize all floor text information, and organize the time series label to obtain the operation and maintenance integration information set.

[0058] Specifically, based on the multi-source splicing sequence obtained previously, first disassemble the text content containing floor numbers and date tags from this sequence, and then compare these text contents with the operation status descriptions registered in the maintenance log one by one. During this process, it is necessary to perform word segmentation and part-of-speech tagging on the valid words in each text record, and list the words with a frequency of occurrence less than 3 times as suspected irrelevant entries and remove them. The frequency threshold of 3 is an empirical value obtained by testing the text parsing accuracy at different frequencies. Immediately afterwards, use a text classification model based on a bidirectional long short-term memory network to identify potential operation instruction features and confirm the event type. The input part of this model includes the serialized word segmentation vectors and the attached floor numbers and date tags. The first layer of long short-term memory units is used to extract context dependencies, and the second layer of long short-term memory units performs deeper screening on the extracted features. The output layer of the model uses the Softmax function to predict various event types. During training, randomly divide 80% of the data from the confirmed sample records for training and 20% for validation. Set the learning rate to 0.001 and update the weight parameters in the network backward after accumulating errors through the cross-entropy loss function. Take 300 iterations to ensure sufficient learning for different types of instruction texts. Stop training when the accuracy of the validation set stabilizes above 95% for 20 consecutive iterations. After training is completed, perform inference on the newly input text through this network. If the probability corresponding to the event category with the highest prediction score is higher than 0.6, it is determined that this record belongs to the corresponding event type. If all are lower than 0.6, mark this record as pending manual confirmation. Then, based on the building control node information, judge the specific operation instruction source for the text records that have been confirmed or automatically determined. Finally, associate all text records with their corresponding event types and output them to obtain the event annotation group.

[0059] According to the event annotation group obtained previously, first read the floor number and operation status description stored in it, and then compare the maintenance details recorded in the maintenance log item by item. If it is found that there are the same floors and similar operation instructions in the same time period, these instructions will be merged into one. Define the recognition standard of "similar operation instructions" as that the floor numbers are the same within the same date and more than 80% of the words are the same after word segmentation of the instruction text. This 80% threshold is obtained by gradually adjusting the ratio on the training data set and observing the aggregation effect of maintenance information. Then, count the possible noise words and perform deletion or replacement processing on them. Noise words refer to texts that cannot correspond to any known operation instructions at all or have no connection with floor actions. Words with an appearance frequency greater than 5 and an overlap degree of 0 with the known instruction set are listed in the noise list. After completing these operations, combine the date tags to check whether the recorded dates conflict with the normal usage range. For example, compare the dates from January 1, 2020 to December 31, 2100. Records outside this date range will be marked as abnormal times and their sources will be further verified. Records that are confirmed not to meet the floor operation conditions will be excluded. After confirming the data for the complete time period, establish a "floor operation entry table" and attach the corresponding operation instruction type to each piece of information. Finally, organize the processed instruction category information to obtain the operation and maintenance dictionary set.

[0060] According to the operation and maintenance dictionary set obtained previously, read the operation instruction types divided by floor and the confirmed key operation descriptions in the dictionary, and then retrieve the corresponding entries from the sensor note field and extract the floor usage status of these entries together. If the description of the floor occupancy degree in the sensor note field for the same time period matches the instruction type in the operation and maintenance dictionary, then link the event content of this time period with the building control node. In this process, a multi-layer convolutional network based on the attention mechanism can be used to retrieve and fuse the text and floor status. The input of this network includes the vectorized key operation description, floor usage status, and instruction type. The convolutional kernel size of the first convolutional layer is set to 3×3, and the convolutional kernel size of the second convolutional layer is set to 5×5. Immediately following the convolutional layer is the calculation of attention weights, and the attention to the key operation segment is highlighted by taking the dot product of the feature matrix and the learnable attention weight vector. During training, the organized operation and maintenance dictionary text and its corresponding floor usage status are selected as paired samples, and the mean squared error is used as the loss function. Calculate the difference between the predicted value and the target value, set the learning rate to 0.0005 and perform 200 rounds of iteration. After training, input the new record into this network for inference. If the proportion of a certain fused feature in the attention output exceeds 0.7, it is regarded as highly correlated with the key operation, otherwise it is recorded as an ordinary event in the daily occupancy state. Finally, re-check and centrally summarize the text information of all floors and the corresponding time series tags to obtain the operation and maintenance fusion information set.

[0061] The steps for obtaining the coupling feature matrix are as follows:

[0062] Based on the operation and maintenance integration information set, perform time period splitting and check the date continuity, fill in the data breakpoints and record the missing intervals, read the floor usage and compare with the external meteorological values, analyze the floor occupancy rate and calculate the temperature difference, and mark the cross fields to obtain the time series difference data;

[0063] Based on the time series difference data, locate the missing intervals and match the floor usage rate classification records, screen the power peak and valley regions and analyze the change range of meteorological values, combine the temperature difference and time period markers and check the record integrity, generate an interval index group, and obtain the load change interval group;

[0064] Based on the load change interval group, perform coupling analysis and match the floor usage records, retrieve the load peak and locate the correlation between temperature, humidity and power changes, extract the time periods and merge the occupancy rate and meteorological data, and mark the load fluctuation characteristics to obtain the coupling feature matrix.

[0065] Specifically, according to the operation and maintenance integration information set obtained previously, first read the records of each floor in the information set and split them in date order. If it is found that the dates or timestamps between two adjacent records are discontinuous, mark them as breakpoint positions, compare the breakpoint positions with the known floor usage to determine whether there are missing record entries. When it is confirmed that there are missing intervals, fill them in according to the interpolation principle. The interpolation process can be estimated based on the average value of adjacent time periods. For example, when there are discontinuities in temperature, humidity or power values, by calculating the compensation value, then extract the occupancy rate range corresponding to each floor from the floor usage. For example, compare the occupancy rate with the 0% to 100% interval. If it exceeds this interval, mark it as an abnormal occupancy rate and submit it for manual review. At the same time, retrieve the external meteorological data and match it with the corresponding dates. For example, compare the temperature with the 0°C to 90°C interval and the humidity with the 0% to 100% interval. Mark the data that is too high or too low as suspicious values and confirm them manually in the same way. After confirming that there is no obvious deviation, calculate the temperature difference inside and outside the floor in turn. The calculation formula is ΔT = T indoor -T outdoor , correspond the temperature differences of all time periods with the corresponding floor occupancy rates one by one. If it is found that the combination of occupancy rate higher than 80% and temperature difference greater than 15°C, add the "cross field" label. The values of 80% and 15°C are determined by performing a regression analysis on the historical floor operation records and referring to industry practice experience. After completing the above operations, summarize the time periods marked with cross fields and other normal time periods to generate the time series difference data.

[0066] Based on the time series difference data obtained previously, first determine the boundaries of the missing intervals by reading the marked cross fields among them. Check whether there are multiple consecutive records with occupancy rate or temperature difference being empty. If so, add default marks after this paragraph and include the start and end times in the missing interval list. Then, identify the usage scenarios of each floor at different time periods according to the floor utilization rate classification records. If the utilization rate of this floor exceeds 90% on the current day and the external meteorological temperature fluctuation is greater than 10°C, it is considered that there is a high-load scenario. The 90% and 10°C are obtained from the statistical analysis of floor user behavior and meteorological changes. For example, in the test data of 30 days, first find the difference distribution between the daily highest and lowest temperatures, and then use the average value as the fluctuation threshold. Then, screen the power peak and valley regions respectively and set the power peak condition above 3000W. This power threshold is 70% of the maximum power during device operation. When it is monitored that the power remains above 3000W for more than 30 minutes, it is determined as a peak. Combine these peak sections with the temperature difference and time period identification, and check the integrity of each record item by item. If there is a missing timestamp or inconsistent floor number, eliminate this item. Finally, pack the qualified intervals into an interval index group to obtain the load change interval group.

[0067] Based on the load change interval group obtained previously, first screen out several time periods with the most frequent peak states and call the actual usage records of the corresponding floors. When comparing the temperature, humidity, and power values during these time periods, if it is found that the floor temperature fluctuates in the range of 15°C to 35°C and the corresponding power is also within the previously set peak determination range, it is determined as a high-load coupling section. On this basis, a set of feedforward artificial neural network is used to analyze the change rates of temperature and humidity. The input of this network includes three parameters: temperature change rate, humidity change rate, and real-time power. The network structure has two hidden layers. The first hidden layer contains 64 neurons, and the second hidden layer contains 32 neurons. The activation function selects ReLU, and the output layer has 1 neuron for numerical regression. During training, 80% of the records are randomly selected from the historical load change data for training. The loss function uses After each iteration, update the weights, and the learning rate is taken as 0.001. Train for 400 rounds to make the network learn relatively stably. After training is completed, input the new high-load coupling section into this network, and output a value for evaluating the correlation degree of temperature, humidity, and power during this time period. If this value exceeds 0.7, it is further determined that the load fluctuation characteristics of this section are obvious. The 0.7 is obtained by comparing the error results between the network prediction and the actual device power curve. Finally, after confirming that all peak sections are evaluated in the above manner, output the coupling feature matrix.

[0068] The steps to obtain the energy consumption deviation parameter are as follows:

[0069] Based on the coupling feature matrix, segment analysis is performed and high-frequency segments are retrieved. Time windows are divided and indoor air velocity and floor lighting start signals are aligned. Change trends and filter blockage records in maintenance logs are recorded. After checking floor numbers and energy consumption records, segment association information is obtained.

[0070] Based on the segmented correlation information, the lighting status of the floor is monitored and the indoor temperature gradient is compared with the lighting on / off records and the temperature and humidity increments. The abnormal sections are extracted and the filter blockage period is checked. The high-frequency operation points of the exhaust equipment are retrieved and the energy consumption peak is quantified to obtain the high-energy consumption node group.

[0071] Based on the high-energy-consuming node group, read the peak period of lighting equipment and match the synchronous information of the heat exchanger, superimpose the temperature and humidity values ​​of the floors and the air flow rate and compare the power fluctuation trajectory, check the floor number matching and divide the fluctuation intensity level, and merge the abnormal peak segments to obtain the energy consumption deviation parameters.

[0072] Specifically, according to the coupling feature matrix obtained above, the records therein are first divided into several sub-segments according to the time sequence and it is detected whether the power fluctuation exceeds the preset threshold value of 3000W. The threshold value is obtained by statistically analyzing the average power of the equipment during normal operation and floating upward by about 20%. The sections with fluctuation values ​​higher than 3000W and lasting for more than 10 minutes are temporarily defined as high-frequency sections. These high-frequency sections are renumbered in chronological order and divided into equal long time windows in turn. The indoor air flow rate values ​​recorded by the flow rate sensors installed on the floors are time-aligned with the start-up signal of the lighting circuit, where the air flow rate is compared with the range of 0m / s to 5m / s. If the short-term flow rate is greater than 5m / s, it is marked as abnormal, and the lighting start-up signal is referenced. The trigger information of the control loop is used to confirm the start time, and then the filter blockage records registered in the maintenance log are compared and searched one by one according to the floor number to see if they overlap with the time interval of the high-frequency section. After finding the qualified entry, it is marked in the corresponding window. When merging the possible multiple records, it is necessary to check whether the floor number, timestamp and power information are consistent. If there are inconsistent records, they are marked and removed from the window. Finally, the merged window is compared with the energy consumption record. If the energy consumption duration is consistent with the fluctuation range of the high-frequency section and the air flow rate or lighting signal belongs to the same floor number, it is determined that the window has corresponding related clues. After sorting out all windows and their corresponding floors and sensor data, the segmented related information is obtained.

[0073] According to the segmented correlation information obtained previously, first confirm the details of the opening and closing of the lighting status on each floor when it appears in the high-frequency paragraph and record whether there is a situation of a large indoor temperature gradient. For example, compare the indoor temperature with the effective range of 0°C to 40°C set. If the temperature increases by more than 5°C within two consecutive acquisitions, it is recorded as a temperature mutation. Then cross-reference these mutations with the lighting opening and closing records. If the lighting is turned on and the temperature mutations occur continuously for multiple times, then mark this section as an "abnormal section". Further check the filter blockage period against this section and see if it occurs within the same floor and the same time range. If it is confirmed that the blockage period coincides with the above abnormal section, then mark this section as a "high-risk paragraph". Then call the operation information of the exhaust equipment and select the records where the fan speed exceeds 3000 revolutions per minute or the power exceeds 3500W as "high-frequency operation points". This 3500W threshold is set in combination with the peak operating power of the exhaust equipment and the actual operating conditions. When quantifying the energy consumption after locating these high-frequency operation points, the energy consumption value for each period can be calculated according to E = P × t. Match the calculation result with the floor number. If the energy consumption value exceeds 10000 joules within one hour, it is regarded as an energy consumption peak. This value is set by analyzing multiple groups of exhaust equipment operation data and referring to its normal working range. After integrating this energy consumption peak information, a high-energy consumption node group is obtained.

[0074] According to the high-energy consumption node group obtained previously, select the records of the peak periods of the lighting equipment on each floor and call the operation data of the heat exchanger within the same time window. Read the floor temperature, humidity, and the indoor air velocity measured by the sensor in these operation data. If the temperature is between 25°C and 35°C, the humidity is between 30% and 70%, and the air velocity is greater than 3m / s, then superimpose this section with the lighting equipment peak and record the power fluctuation trajectory. Then calculate the fluctuation amplitude according to ΔP = P end -P start If ΔP is greater than 800W and the duration exceeds 15 minutes, mark this power fluctuation as "strong". Check this section with a large intensity item by item against the floor number. If they all correspond to the same floor and occur within the same time period, then this section is regarded as a complete abnormal peak section. Finally, when summarizing all abnormal peak sections, if their quantity proportion exceeds 30%, it is considered that the overall fluctuation is frequent. This 30% is obtained by calculating the proportion of the data observed continuously for 30 days and taking the average value. After confirming all abnormal peak sections, classify and merge these sections into a complete index set and extract the part of the record where the difference exceeds the limit. Organize all the peak section information and summarize the energy consumption difference to obtain the energy consumption deviation parameter.

[0075] The steps to obtain the adaptive calibration matrix are as follows:

[0076] Based on the energy consumption deviation parameter, extract the current peak value of floor equipment and retrieve the power fluctuation period. Compare with the indoor temperature and humidity and the operation information of the exhaust equipment, count the deviation data and mark the difference level, aggregate the changes in floor usage and record the corresponding time period to obtain the monitoring and comparison record;

[0077] Based on the monitoring and comparison record, read the input signal of the building control terminal and screen the energy distribution instructions, retrieve the changes in floor usage and check the scheduling period of the exhaust equipment, superimpose the operation instruction records in the maintenance log and analyze the current peak and valley to obtain the scheduling list group;

[0078] Based on the scheduling list group, merge the operation instructions and status weights and analyze the changes in floor usage, verify the energy rationing ratio and record the calibration time on the sensor side, check the response of the building control terminal and summarize the floor temperature and humidity deviation and current fluctuation information to obtain the adaptive calibration matrix.

[0079] Specifically, according to the energy consumption deviation parameter obtained previously, first read the real-time current of the floor equipment and retrieve its peak distribution based on the records per minute. If it is found that the current reaches or exceeds 15A and lasts for more than 5 minutes, it is determined as a peak section. This 15A threshold is obtained by statistically analyzing the normal usage range of multiple sets of equipment. Then, divide the start and end times of each peak section at the minute level and calculate the fluctuation period within a 24-hour cycle. If the time interval between adjacent peak sections is less than 10 minutes and the peak amplitude difference does not exceed 5A, these two sections are merged into a periodic fluctuation section. After the preliminary division of the peak and the fluctuation period, retrieve the recorded indoor temperature and humidity and the operation situation of the exhaust equipment in the floor. Compare the temperature and humidity with the effective ranges of 0°C to 40°C and 0% to 100% respectively. If the temperature or humidity exceeds the effective range, it is marked as abnormal data and confirmed by context interpolation. At the same time, read the operation information of the exhaust equipment and compare the segments where its rotation speed is higher than 3000 revolutions per minute or the power exceeds 3000W. Map these segments to the timeline of the peak section. If it is found that the overlap degree between the high-speed operation of the exhaust equipment and the current peak is higher than 70%, it is considered to have a relatively high correlation. During this process, count the energy consumption deviation values of the current section one by one. For example, use the discrete form of ΔE = ∫(I×V)dt to sum the product of current and voltage over time for estimation. The voltage is calculated at a fixed value of 220V and corrected in combination with the floor electricity consumption characteristics. Finally, divide the statistically obtained deviation data into several levels according to the difference size. If the deviation value of a certain time period is more than 30% higher than the baseline, it is marked as "high difference level". This 30% is obtained by linear regression and mean analysis of historical operation data. Aggregate these marked results with the floor usage one by one and fill them into the corresponding time segments. After summarizing all the records, obtain the monitoring and comparison record.

[0080] According to the monitoring comparison records obtained above, we first retrieve the command signals input by the building control terminal in each time period, and divide them into commands related to energy distribution and other categories of commands. After identifying the commands related to energy distribution, we match the execution time and target floor of the command with the monitoring comparison records. If the occupancy rate of the floor usage changes by more than 50% in the same time period, this section of the command is marked as high-load scheduling. The 50% threshold combines the actual occupancy rules of common office buildings in different time periods. For example, we can obtain the difference between the peak and valley usage of a building on working days and non-working days by counting the difference. Then we read the scheduling information of the exhaust equipment and Check whether the speed or power increases significantly during the execution of the command. If the speed is greater than 3000 rpm or the power is greater than 3500W and lasts for more than 10 minutes, it is determined that the scheduling action has a significant impact. Subsequently, all operation instruction records in the maintenance log are compared with the retrieved current peak and valley data. If the current decreases by more than 10A or increases by more than 10A relative to the peak value of the previous period, it is regarded as an obvious peak-valley change and is marked in the corresponding time period. When the high-load scheduling and peak-valley changes of all floors are confirmed, the results are summarized according to the floor number and time sequence to obtain the scheduling list group.

[0081] According to the dispatch list group obtained above, each instruction and its corresponding state weight value are combined and processed. The state weight value is obtained by analyzing the contribution rate of different instructions to the floor energy consumption level. For example, the instruction to turn on the lighting and exhaust equipment at the same time is given a higher weight of 1.2, the instruction to turn on the lighting or exhaust equipment only is given a medium weight of 1.0, and the weight of the instruction not turning on the above equipment is 0.8. After completing the matching of instructions and weights, combined with the change records of floor usage, each is analyzed one by one. If the occupancy rate rises from 30% to more than 80% during the time period, it is marked as a significant usage change. The threshold is obtained by examining the common floors at different times. The energy consumption scheduling instructions for such periods of large usage changes are checked in detail, and the correspondence between the energy allocation ratio and the total equipment power is verified. If the allocation ratio exceeds 0.9 and the actual equipment power is still insufficient to meet the statistical energy demand, a calibration time is immediately recorded on the sensor side. The sensor side calibration time is intended to mark the power gap of the current floor. After the next scheduling is completed, the power consumption curve before and after this time point is compared, and the floor temperature and humidity deviation and current fluctuation information are further checked when the building control end responds. The results of all the above verification processes are centrally organized to obtain the adaptive calibration matrix.

[0082] The steps for obtaining the energy consumption prediction sequence are as follows:

[0083] Based on the adaptive calibration matrix, accumulate the temperature and humidity sampling differences of each floor and verify the date consistency, select the meteorological values and wind speed elements and count the corresponding distributions, mark the usage of each floor and the switching periods of the indoor operating states, merge the maintenance logs and compare the operation information of the exhaust equipment to obtain the iterative difference records;

[0084] Based on the iterative difference records, analyze the coupling relationship between the floor usage and meteorological values and estimate the rate of change of temperature and humidity, retrieve the peak and valley sections of the current and compare the energy rationing ratios, count the operating period distributions of each floor and then mark the abnormal fluctuation information to obtain the fluctuation curve groups;

[0085] Based on the fluctuation curve groups, read the floor occupancy rate and meteorological data and proofread the temperature and humidity differences, verify the exhaust equipment statistics and summarize the node energy consumption changes in the control and scheduling list and then analyze the floor power trend, screen the key time periods and draw the time series trend chart to obtain the energy consumption prediction sequences.

[0086] Specifically, according to the adaptive calibration matrix obtained above, the cumulative difference of temperature and humidity of each floor is first calculated at the set sampling frequency of 5 minutes. When the temperature difference between two adjacent sampling points exceeds 8°C or the humidity difference exceeds 15%, it is marked as a "high difference entry". The 8°C and 15% are set based on the comparative analysis results of the historical operation records of the floor and combined with industry experience. Then check whether the date and time stamps of each entry are continuous. If a date is not in the recognized range (such as January 1, 2020 to December 31, 2100) or the timestamp jumps for more than 30 minutes, it is regarded as an abnormal section. These abnormal sections are listed separately and linear interpolation is performed on the temperature and humidity before and after. In order to more accurately judge the impact of floor usage and external meteorological values, it is also necessary to read the temperature, relative humidity and wind speed provided by external meteorological monitoring equipment. The temperature is compared with 0°C to 50°C, the humidity is compared with 0% to 100%, and the wind speed is compared with 0m / s to 20m / s. If any item exceeds the valid range, it is marked as The suspected observation deviation is detected and the number of the floor is recorded. Then, the operation content of the date is selected from the building maintenance log and combined with the start and stop information of the exhaust equipment for parallel verification. If the exhaust equipment is turned on and the power exceeds 2000W in the same time period, it is marked as "high-intensity exhaust". If the power is between 1000W and 2000W, it is marked as "medium-intensity exhaust". If it is less than 1000W, it is marked as "low-intensity exhaust". These intervals are set after numerical statistics of the exhaust equipment specification and long-term operation data. Then, the occupancy rate (0% to 100%) and indoor operating status (whether the lighting is turned on, whether the air conditioner is enabled, etc.) in the floor usage are combined to identify the switching period. If a floor jumps rapidly from an occupancy rate of less than 10% to more than 60% within 10 minutes, it is recorded as a "fast switching section". After confirming the above information, the marked high-difference items, suspected observation deviation items, exhaust equipment intensity information and floor switching periods are merged together. By integrating these elements and calculating the cumulative temperature and humidity differences in adjacent time periods, the iterative difference records are obtained.

[0087] According to the iterative difference records obtained previously, we first compare them with the corresponding meteorological values ​​according to the floor division, and perform coupling relationship analysis on the change amplitude of temperature and humidity in the effective range. When the floor usage (such as the occupancy rate increases from 30% to more than 70%) and the external temperature fluctuation (such as from 5°C to 15°C) show a synchronous growth trend, and there is a rising section for more than 10 minutes on the cumulative curve of the indoor temperature or humidity difference, the section is regarded as a "high coupling zone". This judgment is based on the results of the floor historical operation statistics and multiple regression analyses. When the occupancy rate and meteorological temperature increase at the same time, a two-layer feedforward network is used to estimate the temperature and humidity change rate. The input of this network includes the indoor and outdoor temperature difference, the indoor and outdoor humidity difference and the occupancy rate. Each hidden layer contains 64 neurons, and the activation function uses ReLU. The output is the estimated temperature and humidity change rate value. During training, the floor operation data of the past 30 days are collected as the training set, and the mean square error is used. to measure the difference between the predicted value and the true value, the learning rate is set to 0.001, and after 400 rounds of training, the stable error convergence is achieved in the validation set. After the training is completed, the new record is sent to the network for inference and the estimated value of the temperature and humidity change rate is obtained. If the actual record differs from the estimated value by more than 30%, it is marked as an abnormal coupling segment. At the same time, the content related to the current peak and valley in the iterative difference record is retrieved. When it is detected that the energy allocation of a certain floor exceeds 0.8 ratio (the 0.8 is obtained by comparing the maximum load of the equipment with the normal energy supply), and there is a significant peak-to-valley change in this period (such as a large jump in the current between 5A and 20A), the usage period distribution of the floor throughout the day is counted, and "abnormal fluctuation information" is added for this section. When the data of all floors are marked as above, a fluctuation curve group is obtained.

[0088] According to the obtained fluctuation curve group, first read the occupancy rate and meteorological data of each floor and align them on the same time coordinate. Then, check item by item whether the difference between the indoor temperature and the outdoor temperature exceeds 15°C. If it exceeds, mark it as the "high temperature difference section". At the same time, conduct further verification according to the previously statistically obtained rotational speed and power information of the exhaust equipment. If the rotational speed of the exhaust equipment exceeds 3000 revolutions per minute or the power is greater than 2500W during the same period (the thresholds of 3000 revolutions and 2500W are determined by checking the equipment technical specifications and actual long-term monitoring data), mark this period as "high ventilation load" additionally. Then, retrieve all operation instruction nodes in the control and scheduling list and check the energy consumption changes of these nodes on the current floor. For example, when a certain scheduling instruction requires the floor to use the air conditioning mode and the occupancy rate soars from 30% to 90%, it is necessary to analyze the power trajectory. If the power climbs from 500W to 2000W within 5 minutes, record a significant power increase. Immediately afterwards, align the significant power increase sections of all floors in chronological order and draw the time series distribution. In this time series distribution, screen the floors with multiple sudden increases within 15 minutes and summarize them into key time periods. These key time periods are usually from 8:00 to 10:00 in the morning or from 1:00 to 3:00 in the afternoon on weekdays. Finally, mark all key time periods in the curve and form an overview of the power trend of this building over several days, output the drawn overall trend chart and collate the final estimated results uniformly to obtain the energy consumption prediction sequence.

[0089] The above is only a preferred embodiment of the present invention, and does not impose other forms of limitations on the present invention. Any person skilled in the art may use the disclosed technical content to make changes or modifications into equivalent embodiments with equivalent changes and apply them to other fields. However, as long as it does not depart from the technical solution content of the present invention, any simple modification, equivalent change and modification made to the above embodiments based on the technical essence of the present invention still fall within the protection scope of the technical solution of the present invention.

Claims

1. An artificial intelligence-based building energy consumption prediction method, characterized in that It includes the following steps: When performing data cleaning and time period division based on floor usage and sensor monitoring values, duplicate entries are removed, abnormal records are eliminated, and the time dimension is reconstructed. When screening temperature, humidity, and power information, the measurement units are unified and the timestamps are calibrated. The operation records of the heat exchanger and the maintenance items of the exhaust equipment in the maintenance log are associated to obtain a multi-source splicing sequence; Based on the multi-source splicing sequence, text tokenization and operation event annotation are performed, the date range is delimited, noise words are filtered, the building control nodes in the knowledge graph are linked, the floor numbers and operation instruction types are identified, and the operation instructions and sensor note records in the maintenance log are integrated to obtain an operation and maintenance integration information set; Based on the operation and maintenance integration information set, time series splitting and difference calculation are performed, missing intervals are detected, breakpoint data is filled, the temperature difference is cross-compared when matching the floor usage and external meteorological indicators, the usage rate classification records in the operation archives are summarized, and the load change intervals are screened to obtain a coupling feature matrix; Based on the coupling feature matrix, segment analysis and time period association matching are performed, high-frequency segments are split, discontinuous time periods are merged, the change trends are recorded when monitoring the indoor air circulation rate and the floor lighting start signal, the filter blockage records in the maintenance log are verified, the energy-consuming nodes are identified, and the fluctuation intensity is quantified to obtain an energy consumption deviation parameter.

2. The artificial intelligence-based building energy consumption prediction method according to claim 1, wherein, It also includes: Based on the energy consumption deviation parameter, real-time monitoring data comparison and difference push are performed, the current fluctuations of the floor equipment are extracted and the peaks are recorded. When adjusting the input signal of the building control end and correcting the energy distribution ratio, the changes in floor usage are compared, the operation period information in the maintenance log is merged, and after outputting the control and scheduling list, the status weights are integrated to obtain an adaptive calibration matrix; Based on the adaptive calibration matrix, the sampling differences of the floor temperature and humidity are accumulated, the floor usage and meteorological data are fused, the distribution characteristics are statistically analyzed, the wind speed elements are synchronously recorded, the energy consumption trends of each node in the control and scheduling list are analyzed, and the predicted time series fluctuation curve is output to obtain an energy consumption prediction sequence.

3. The method for predicting building energy consumption based on artificial intelligence according to claim 1, wherein The steps for obtaining the multi-source splicing sequence are as follows: Based on the floor usage and sensor monitoring values, when performing record verification and duplicate entry investigation, the same records are removed and damaged data is removed, the time series is reorganized and the monitoring time periods are marked, abnormal dates are excluded, and the floor numbers are synchronized and aligned to obtain time period corrected data; Based on the time period corrected data, when screening the temperature, humidity, and power fields, the measurement units are unified and the sampling frequencies are aligned, invalid measurement points are removed, and the floor positions are marked. The operation time periods of the heat exchanger in the maintenance log are associated and compared with the exhaust equipment maintenance records to obtain a screened temperature, humidity, and power group; Based on the screened temperature, humidity, and power group, missing time period verification and operation records are merged, the time series index is reconstructed, the floor coding fields are allocated, the states of adjacent time periods are superimposed, and the overall mapping of the building is generated. Duplicate sampling identifiers are checked, and the collaborative operation information of the heat exchanger and the exhaust equipment is summarized to obtain a multi-source splicing sequence.

4. The method for predicting building energy consumption based on artificial intelligence according to claim 1, wherein The steps for obtaining the operation and maintenance integration information set are as follows: Based on the multi-source splicing sequence, extract the floor number and date label when performing text parsing and event splitting, filter out irrelevant entries and match the maintenance log operation status description, associate the building management and control node information and identify the source of the operation instruction to obtain the event annotation group; Based on the event annotation group, perform floor field comparison and map maintenance log record inspection, remove duplicate instructions and merge similar records in the same period, identify noise words and establish floor operation entry table, proofread date tags and associate operation instruction types, and obtain operation and maintenance dictionary set; Based on the operation and maintenance dictionary set, the command sources are classified and the sensor remark fields are matched, the floor usage status and key operation descriptions are split, the building management and control nodes are linked and the operation event content is merged, all floor text information is summarized and the time series labels are sorted to obtain the operation and maintenance fusion information set.

5. The artificial intelligence-based building energy consumption prediction method according to claim 1, wherein, The steps of obtaining the coupling characteristic matrix are: Based on the operation and maintenance fusion information set, split the time period and check the date continuity, fill in the data breakpoints and record the missing intervals, read the floor usage and compare the external meteorological values, analyze the floor occupancy rate and calculate the temperature difference, mark the cross field and obtain the time series difference data; Based on the time series difference data, locate the missing intervals and match the floor usage classification records, screen the power peak and valley areas and analyze the amplitude of meteorological value changes, combine the temperature difference with the time period mark and verify the integrity of the records, generate the interval index group, and obtain the load change interval group; Based on the load change interval group, coupling analysis is performed and floor usage records are matched, load peaks are retrieved and the association between temperature, humidity and power changes is located, time periods are extracted and occupancy and meteorological data are merged, and the coupling feature matrix is ​​obtained after the load fluctuation characteristics are marked.

6. The building energy consumption prediction method based on artificial intelligence according to claim 1, wherein, The steps for obtaining the energy consumption deviation parameter are: Based on the coupling feature matrix, segment analysis is performed and high-frequency segments are retrieved, time windows are divided and indoor air velocity and floor lighting start signals are aligned, change trends and filter blockage records in maintenance logs are recorded, and segment association information is obtained after checking floor numbers and energy consumption records; Based on the segmented association information, the lighting status of the floor is monitored and the indoor temperature gradient is compared with the lighting on / off records and the temperature and humidity increments, the abnormal sections are extracted and the filter blockage period is checked, the high-frequency operation points of the exhaust equipment are retrieved and the energy consumption peak is quantified to obtain the high-energy consumption node group; Based on the high-energy consumption node group, read the peak period of the lighting equipment and match the synchronization information of the heat exchanger, superimpose the temperature and humidity values ​​of the floors and the air flow rate and compare the power fluctuation trajectory, check the floor number matching and divide the fluctuation intensity level, and obtain the energy consumption deviation parameter after merging the abnormal peak segments.

7. The method for predicting building energy consumption based on artificial intelligence according to claim 2, wherein The steps of obtaining the adaptive calibration matrix are: Based on the energy consumption deviation parameter, extract the peak current of floor equipment and retrieve the power fluctuation period, compare the indoor temperature and humidity and exhaust equipment operation information, count the deviation data and mark the difference level, aggregate the changes in floor usage and record the corresponding time period to obtain the monitoring comparison record; Based on the monitoring and comparison records, read the input signals of the building control terminal and filter the energy distribution instructions, retrieve the changes in floor usage conditions and check the scheduling periods of the exhaust equipment, superimpose the operation instruction records in the maintenance log and analyze the current peak and valley, to obtain a scheduling list group; Based on the scheduling list group, merge the operation instructions and status weights and analyze the changes in floor usage conditions, verify the energy rationing ratio and record the calibration time on the sensor side, check the response of the building control terminal and summarize the floor temperature and humidity deviation and current fluctuation information, to obtain an adaptive calibration matrix.

8. The method for predicting building energy consumption based on artificial intelligence according to claim 2, characterized in that The steps for obtaining the energy consumption prediction sequence are as follows: Based on the adaptive calibration matrix, accumulate the sampling differences of floor temperature and humidity and verify the date consistency, select the meteorological values and wind speed elements and count the corresponding distributions, mark the switching periods of floor usage conditions and indoor operation states, merge the maintenance log and compare the operation information of the exhaust equipment, to obtain an iterative difference record; Based on the iterative difference record, analyze the coupling relationship between floor usage conditions and meteorological values and estimate the rate of change of temperature and humidity, retrieve the current peak and valley sections and compare the energy rationing ratio, count the operating period distributions of each floor and then mark the abnormal fluctuation information, to obtain a set of fluctuation curves; Based on the set of fluctuation curves, read the floor occupancy rate and meteorological data and proofread the temperature and humidity differences, verify the exhaust equipment statistics and summarize the node energy consumption changes in the control and scheduling list and then analyze the floor power trend, select the key time periods and draw a time series trend chart, to obtain the energy consumption prediction sequence.

Images