Dormitory hot water consumption prediction method and system fusing student work and rest and weather

By using a CNN-LSTM model to integrate student schedules with multi-dimensional weather data, the problem of supply-demand mismatch and energy waste in hot water supply management in university dormitories has been solved, achieving accurate water demand prediction and improved management efficiency.

CN121920765APending Publication Date: 2026-04-24GUANGZHOU UNIVERSITY
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202610041383.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-01-13
Publication Date
2026-04-24

Smart Images

  • Figure CN121920765A_ABST
    Figure CN121920765A_ABST
Patent Text Reader

Abstract

The invention discloses a dormitory hot water consumption prediction method and system fusing student work and rest and weather, and the method comprises the steps: firstly obtaining historical hot water consumption time sequence data, and student work and rest data and weather data which are used for representing the future prediction time period state; then, a work and rest intensity feature vector, a weather feature vector and a hot water consumption time sequence feature vector are constructed and fused, and then the fused feature vectors are input into a trained CNN-LSTM prediction model to output a future water consumption prediction value. Wherein the model adopts a depth separable convolution structure and introduces a time sequence attention module to enhance the attention to a water consumption peak period, and the method further comprises a model training step and a feedback optimization closed loop based on actual consumption data, so that the water consumption rule of the college dormitory is accurately captured. On-demand scheduling of hot water and energy conservation and consumption reduction of the system are realized.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of smart campus energy management technology, and more specifically, to a method and system for predicting dormitory hot water consumption by integrating student schedules and weather. Background Technology

[0002] With the rapid development of higher education and the continuous expansion of campuses in my country, the supply and demand balance and energy efficiency of hot water supply in student dormitories, as a basic public service, have become an increasingly important focus. Currently, most university dormitories still employ a crude "experience-based scheduling" model for hot water supply management, relying on historical data. Logistics staff manually adjust supply strategies based on macro-level factors such as seasons and semester stages. This model struggles to accurately meet the highly dynamic and concentrated water demands of students, generally leading to two major problems: first, a mismatch between supply and demand, resulting in insufficient hot water supply during peak hours (such as after evening self-study), affecting student experience, while excessive supply during off-peak hours (such as class time) often leads to energy waste; second, overall energy consumption is relatively high.

[0003] To improve management efficiency, relevant technical fields have attempted to introduce predictive methods. Existing technical solutions mainly fall into two categories: one is prediction methods based on traditional statistical models, such as time series analysis models using the Autoregressive Integral Moving Average (ARIMA) model, which primarily rely on historical consumption data for extrapolation. The other is prediction methods based on shallow machine learning, such as using backpropagation (BP) neural networks, which use historical consumption and single or limited factors like ambient temperature as input features. However, applying these methods to the unique scenario of university dormitories presents significant limitations: First, these methods fail to fully consider the core characteristics of university dormitories—high population density and highly uniform schedules—ignoring the decisive influence of course schedules, holidays, exam weeks, and dormitory return patterns on students' water usage behavior; second, their consideration of weather factors is rather one-sided, often using only ambient temperature as the sole indicator, neglecting the combined effects of multi-dimensional weather conditions such as humidity, wind speed, and precipitation; finally, traditional models are insufficient in capturing complex spatiotemporal dependencies, making it difficult to model the dynamic correlation of water usage between dormitory areas and the long-term periodic changes in water demand. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for predicting dormitory hot water consumption by integrating student schedules and weather, in order to solve the above-mentioned problems existing in the prior art.

[0005] The application is as follows: In a first aspect, the present invention provides a method for predicting dormitory hot water consumption by integrating student schedules and weather conditions, comprising the following steps: S1. Obtain the data required for prediction, including: hot water consumption time series data of the target dormitory within a preset time window before the prediction time, as well as student daily routine data and weather data used to characterize the state of the future prediction period; S2. Extract and quantify features from the student daily routine data to construct a daily routine intensity feature vector; preprocess the weather data to construct a weather feature vector; preprocess the hot water consumption time series data to extract its hot water consumption time series features; and fuse the hot water consumption time series features, daily routine intensity feature vector, and weather feature vector to generate a fused feature vector. S3. Input the fused feature vector into the trained CNN-LSTM prediction model and output the predicted value of hot water consumption for a specified future period.

[0006] Furthermore, in step S1, the student schedule association data includes time attribute data and behavior association data; the time attribute data includes at least the information on holidays, weekends and exam weeks stipulated in the school calendar; the behavior association data includes at least the dormitory access control system's return-to-dormitory records and the timetable data corresponding to the future predicted period; the weather data includes temperature, humidity, wind speed and precipitation data.

[0007] Furthermore, in step S2, the construction of the activity intensity feature vector includes: calculating the estimated in-dormancy rate corresponding to the future prediction period based on the dormitory return record; determining whether the future prediction period is in the concentrated water use sensitive period after the end of the course based on the time schedule data, and generating corresponding identification features; and generating time identification features representing holidays, weekends, and exam weeks based on the time attribute data. The estimated sleep rate, the water-sensitive period marker, and the time marker are concatenated, and the concatenated feature vector is standardized to form the sleep intensity feature vector.

[0008] Furthermore, in step S2, the extraction of hot water consumption time-series characteristics includes extracting its periodic characteristics and trend characteristics; the periodic characteristics include at least a fluctuation pattern encoding with a period of days or weeks.

[0009] Furthermore, in step S3, the CNN model of the CNN-LSTM prediction model adopts a depthwise separable convolutional structure to efficiently extract local patterns related to water use behavior from the fused feature vector.

[0010] Furthermore, the CNN-LSTM prediction model also includes a temporal attention module; the temporal attention module is used to assign dynamic weights to the input temporal feature sequence to enhance the model's attention to peak water usage periods, and the temporal feature sequence is composed of the fused feature vectors of multiple time steps in chronological order.

[0011] Furthermore, the workflow of the temporal attention module includes: S71. Receive the time-series feature sequence as input; S72. Calculate the attention score for each time step in the temporal feature sequence; S73. Based on the attention scores, the features at each time step are weighted and summed to generate a context vector; S74. Fuse the context vector with the temporal feature sequence to output a weighted temporal feature sequence.

[0012] Furthermore, it also includes the model training steps: S81. Collect time-series data on hot water consumption, student daily routine data, and weather data of the target dormitory during historical time periods, and divide them into training set and validation set, and construct training samples and validation samples according to the method described in claims 1 to 7. S82. Using the training samples, with mean squared error or mean absolute percentage error as the loss function, use the Adam optimizer to iteratively train the CNN-LSTM prediction model; S83. During iterative training, monitor model performance based on the validation samples, and save the current optimal model when the model loss function converges; S84. Using the mean absolute percentage error as the evaluation index, calculate the prediction accuracy of the current best model on the validation set; if the prediction accuracy is not lower than a preset threshold, then complete the training and determine the final model; otherwise, after optimizing the feature construction strategy, repeat step S82.

[0013] Furthermore, it also includes: S4. Application steps: Convert the predicted hot water consumption for the specified future period into control parameters or scheduling suggestions for the dormitory hot water supply system, and output them to the logistics management platform; S5. Feedback optimization step: Collect actual hot water consumption data, compare it with the predicted value, and use the difference data to update the training dataset of the CNN-LSTM prediction model to achieve iterative optimization of the model.

[0014] Secondly, the present invention provides a dormitory hot water consumption prediction system that integrates student schedules and weather conditions, the system comprising: The data acquisition and preprocessing module is used to acquire and process hot water consumption time-series data, student daily routine data, and weather data; The feature engineering module is used to construct and fuse the activity intensity feature vector, weather feature vector, and hot water consumption time series features to generate a fused feature vector. The prediction model module has a built-in CNN-LSTM prediction model trained by the steps described in claim 8, which is used to receive the fused feature vector and output the prediction value. The results output and scheduling suggestion module is used to convert the predicted values ​​into scheduling instructions for hot water supply equipment. The model optimization module is used to collect actual hot water consumption data, compare it with the predicted value, and trigger strategy optimization of the feature engineering module or retraining of the prediction model module based on the comparison result.

[0015] Compared with the prior art, the embodiments of the present invention achieve the following beneficial effects: This invention constructs a feature system that comprehensively characterizes water usage drivers by deeply integrating student daily routine data (such as class schedules, holidays, and dormitory return rates) with multi-dimensional weather data. Combined with an optimized CNN-LSTM model, particularly the introduced temporal attention mechanism, the model can accurately capture the unique, routine-driven "peak-hour" water usage patterns found only in university dormitories. This method improves the prediction accuracy of peak water usage periods, enabling logistics departments to schedule hot water supply equipment in advance and precisely, avoiding hot water shortages during peak hours and energy waste during off-peak hours, significantly improving students' water usage experience and service satisfaction.

[0016] This invention is specifically designed for the unique scenario of university dormitories, characterized by high population density, regular work and rest schedules, and highly concentrated needs. The design of the model's input features is directly derived from campus management practices. Furthermore, the use of lightweight model structures such as depthwise separable convolution significantly reduces computational complexity while maintaining performance, making the model easy to deploy and run on existing campus servers or edge computing devices, thus enhancing the practicality and scalability of the technology.

[0017] This invention continuously collects actual water usage data and compares it with predicted values, enabling the system to automatically monitor model performance degradation. When performance falls short, the system can trigger feature engineering optimization or model retraining, allowing the entire prediction system to adaptively track slow dynamics such as long-term changes and seasonal variations in students' daily routines. This self-evolutionary capability significantly reduces the system's later maintenance costs and reliance on manual intervention, ensuring the reliability and accuracy of the prediction system in long-term operation. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating a method for predicting dormitory hot water consumption that integrates student schedules and weather, as provided in an embodiment of the present invention. Detailed Implementation

[0019] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments given herein are for illustration and explanation only and are not intended to limit the present invention.

[0020] It should be noted that many specific details are set forth in the following description in order to provide a full understanding of the present invention. However, the present invention may have other embodiments. Therefore, the scope of protection of the present invention is not limited to the specific embodiments disclosed below.

[0021] Example 1

[0022] See attached document Figure 1 As shown in the figure, this embodiment provides a method for predicting dormitory hot water consumption by integrating student schedules and weather, which includes the following steps: S1. Obtain the data required for prediction, including: hot water consumption time series data of the target dormitory within a preset time window before the prediction time, as well as student activity data and weather data to characterize the state of the future prediction period.

[0023] The hot water consumption time series data reflects the historical sequence of past water use behavior. In this embodiment, the hot water consumption time series data is collected and uploaded to the data center at fixed time intervals (such as 5 minutes or 1 hour) by smart water meters or flow sensors deployed on the dormitory water supply pipeline.

[0024] The student schedule data includes time attribute data and behavior-related data. The time attribute data is obtained directly from the school's academic affairs system or academic calendar, and includes at least information on holidays, weekends, and exam weeks as specified in the academic calendar. This information is predetermined and corresponds completely to the future predicted period. The behavior-related data includes at least dormitory return records from the dormitory access control system and timetable data corresponding to the future predicted period. The timetable data is obtained from the school's course management system and is used to determine whether the future predicted period falls within the "critical water usage period after the end of classes."

[0025] The weather data is obtained by calling authoritative meteorological service APIs (such as the China Meteorological Administration and Xinzhi Weather) to obtain forecast values ​​of temperature, humidity, wind speed and precipitation probability for the area in the future forecast period.

[0026] All data undergoes cleaning operations after collection, including aligning timestamps, handling missing values, and removing obvious outliers, to ensure data quality.

[0027] S2. Perform feature extraction and quantification on the student's daily routine data to construct a daily routine intensity feature vector; preprocess the weather data to construct a weather feature vector; preprocess the hot water consumption time series data to extract its hot water consumption time series features; and fuse the hot water consumption time series features, daily routine intensity feature vector, and weather feature vector to generate a fused feature vector.

[0028] S21 Constructing a Daily Routine Intensity Feature Vector: This involves quantifying student daily routine data, which characterizes the future behavioral patterns of a student group, into a comprehensive numerical feature vector. The specific construction process is as follows: Estimated Dormitory Occupancy Rate Calculation: The system uses historical and real-time dormitory access records. Based on historical data, it analyzes the average dormitory occupancy rate during historical periods with the same attributes as the predicted future period (e.g., the same day of the week, whether it's a holiday). Simultaneously, it dynamically adjusts the estimate by incorporating real-time dormitory occupancy data. The final calculated estimated dormitory occupancy rate is between 0 and 1, representing the proportion of students expected to remain in the dormitory during the predicted future period.

[0029] Water Concentration Sensitive Period Identifier: The system accesses the timetable data to determine whether the predicted future period falls within a preset time window (e.g., 1 to 3 hours) after the end of the main daily courses. If so, the system identifies this period as a "water concentration sensitive period" where water demand may surge, and generates a binary identifier feature (assigned a value of 1); otherwise, it assigns a value of 0.

[0030] Time attribute identifier: Based on the school calendar data, the system generates the following three independent binary features to characterize the special date attributes of the future prediction period: Whether it is a holiday: If the time period falls on a statutory or school-designated holiday, mark it as 1; otherwise, mark it as 0. Weekend status: Mark 1 if the time period is on Saturday or Sunday, otherwise mark 0; Is it an exam week? If the time period falls within the school's designated exam week, mark it as 1; otherwise, mark it as 0.

[0031] Feature standardization and vector formation: The five feature values ​​generated in the above steps—the estimated sleep rate (continuous value), the concentrated water use sensitivity period identifier (binary), and the time attribute identifier (three binary values)—are concatenated to form a one-dimensional feature array. Subsequently, this array is normalized (e.g., Min-Max standardization) to eliminate scale differences between different features, ultimately outputting a unified sleep intensity feature vector.

[0032] S22 Constructs Weather Feature Vectors: The system acquires forecast values ​​for temperature, humidity, wind speed, and precipitation for the future prediction period. Each meteorological indicator is standardized to eliminate the influence of different dimensions and transform each indicator to a similar numerical range. The processed indicator values ​​are arranged in order to form a weather feature vector, which serves as the conditional feature input characterizing the future environmental state.

[0033] S23 Extracts Temporal Features of Hot Water Consumption: This feature extraction method extracts characteristics that represent the inherent patterns of hot water consumption from historical time-series data to capture temporal patterns in water usage behavior. The system processes the hot water consumption sequence within a preset time window preceding the prediction period, primarily extracting two types of features: Periodicity Features: To capture the recurring patterns in water usage behavior that are cyclical in "days" and "weeks," the system encodes the time point information in the time series data. This involves transforming the intraday time (hour) and intraweek day sequence (day of the week) corresponding to each time point into a set of continuous value features. Periodic encoding is performed using sine and cosine functions, mapping discrete cyclical time variables to a continuous two-dimensional space, enabling the model to understand characteristics such as the proximity of 22:00 and 23:00 and the periodic regression of Mondays.

[0034] Trend / Statistical Characteristics: To characterize the recent level and dynamic changes in water consumption, the system calculates a series of statistics from the time window series. These typically include, but are not limited to: the mean and standard deviation of the series (reflecting the overall level and fluctuations), simple moving average (smoothing noise and highlighting trends), and first difference (characterizing the changes between adjacent time periods).

[0035] The extracted periodic features are combined with trend / statistical features to form a hot water consumption time series feature vector, which is used to comprehensively describe the time series status of historical water use behavior.

[0036] S24 concatenates the activity intensity feature vector, weather feature vector, and hot water consumption time-series feature vector along the feature dimensions. Specifically, the activity intensity feature vector has the following dimensions: The dimension of the weather feature vector is The dimension of the hot water consumption time series feature vector is The dimension of the fused feature vector generated after concatenation is... It retains all the information of each original feature vector and juxtaposes them structurally to form a high-dimensional feature representation that comprehensively represents future behavioral patterns, environmental conditions, and historical time series patterns.

[0037] S3. Input the fused feature vector into the trained CNN-LSTM prediction model and output the predicted value of hot water consumption for a specified future period.

[0038] Temporal Feature Sequence Construction: To capture the dynamic changes in water usage behavior, the current fused feature vector obtained in the above steps is combined and aligned with the historical fused feature vectors from multiple consecutive time steps in chronological order. These sequentially arranged fused feature vectors together constitute a temporal feature sequence. This sequence is a three-dimensional tensor, whose dimensions represent the batch size, time step length, and feature dimension of each time step, thus fully characterizing the evolution of all key factors affecting hot water consumption (student schedules, weather, historical water usage) over a recent continuous period.

[0039] Convolutional Feature Extraction: The aforementioned temporal feature sequence is input into the CNN module of the CNN-LSTM prediction model. This module employs a depthwise separable convolutional structure for processing. This structure first extracts spatial features independently for each channel of the input data through deep convolutional layers, and then fuses the information between channels through pointwise convolutional layers. This effectively extracts local spatiotemporal patterns (such as the correlation between water use behavior and specific weather and work / rest conditions over a short period of time) while significantly reducing the number of model parameters and computational complexity.

[0040] Temporal attention weighting: The temporal feature sequence output by the convolutional module is fed into the temporal attention module. The workflow of this module includes the following steps: S71 receives input: The temporal attention module receives the temporal feature sequence as input.

[0041] S72 calculates the attention score: The module automatically calculates the importance weight of each historical time step in the original time series feature sequence for the current prediction, i.e., the attention score, through internal learnable parameters; S73 generates a context vector: using the calculated attention score, the features of all time steps in the original temporal feature sequence are weighted and summed to synthesize a context vector that gathers global key information; S74 Fusion and Output: This process fuses the context vector with the original temporal feature sequence (e.g., by element-wise addition or channel concatenation) to output a recalibrated temporal feature sequence. This process enhances features associated with key patterns such as water usage peaks and suppresses non-critical or redundant information.

[0042] Temporal Dependency Modeling and Predictive Output: The attention-weighted temporal feature sequence is ultimately input into the Long Short-Term Memory (LSTM) network module. The LSTM module, through its internal forgetting gate, input gate, and output gate, learns and models the complex long-term and short-term temporal dependencies in water demand, such as daily periodic fluctuations, weekend effects, and the persistent impact of holidays. The hidden state at the last time step of the LSTM module, after a linear transformation through a fully connected layer, directly outputs a scalar value, which is the model's final prediction of dormitory hot water consumption for a specified future period.

[0043] S4 Application Steps: Transform the numerical results predicted by the CNN-LSTM prediction model into control commands that can directly drive the operation of the physical system, completing the closed loop from data analysis to actual control.

[0044] The CNN-LSTM prediction model forecasts the hot water consumption for a specified future period. Based on this forecast, corresponding control parameters for the hot water supply system or specific scheduling suggestions are generated through preset mapping rules or control algorithms.

[0045] Control parameters are typically generated based on predicted values, dynamically setting the target temperature, start / stop time, or operating power of the circulating pump for the hot water heating equipment. Scheduling recommendations may include operational instructions for support staff, such as "Hot water consumption is expected to increase by 15% during the evening peak; it is recommended to start the standby heater 30 minutes in advance."

[0046] Ultimately, the generated control parameters or scheduling suggestions are pushed to the campus logistics intelligent management platform in real time. This platform can automatically execute control commands or provide decision support for managers, thereby enabling on-demand scheduling and precise control of the hot water supply system.

[0047] S5 Feedback Optimization Steps: Establish a continuous learning mechanism for the CNN-LSTM prediction model, and use actual running data to iteratively optimize the prediction model in order to improve its long-term adaptability and prediction accuracy.

[0048] After each predicted future period ends, the system collects data on the actual hot water consumption during that period using smart water meters or flow sensors deployed on the dormitory water supply network. Subsequently, the system automatically and precisely compares the actual consumption with the predicted value output by the model for the corresponding period, and calculates the prediction error (such as absolute error, relative error, or mean square error).

[0049] The system archives the actual hot water consumption data with precise timestamps, the corresponding raw input data (student schedule data, weather data), and the calculated prediction error as a complete new data sample into the historical database. This process enables dynamic expansion of the training dataset, ensuring that the dataset reflects the latest student water usage patterns and external environmental conditions.

[0050] The newly accumulated data samples will be used in the next model optimization cycle set by the system (e.g., every Sunday morning). The system will trigger iterative model optimization in one of the following two ways, based on a preset strategy: Incremental learning / fine-tuning: New samples are directly added to the existing training set to perform an incremental training on the deployed CNN-LSTM prediction model, quickly adapting to the latest changes.

[0051] Periodic full training: When new data accumulates to a certain scale (or reaches a fixed period), the system automatically starts the complete model retraining process, using the training set containing all new and old data to generate a brand new, higher-performing prediction model version for subsequent deployment.

[0052] Through continuous feedback optimization steps, the CNN-LSTM prediction model can have the ability to self-evolve, effectively tracking the long-term changes in students' daily routines and slow changes such as seasonal changes, thereby ensuring the high accuracy and high reliability of the CNN-LSTM prediction model throughout its entire life cycle.

[0053] Example 2

[0054] This embodiment details the entire process of offline training, validation, and evaluation of a CNN-LSTM prediction model. This process is typically performed before the system's initial deployment or during regular comprehensive updates.

[0055] S81. The system collects time-series data on hot water consumption in the target dormitory, student daily routine data, and weather data over historical periods. To ensure the fairness of model evaluation, the overall dataset will be divided into two mutually exclusive subsets: a training set and a validation set, either chronologically or randomly. In this embodiment, 70% of the subset can be used as the training set, and the remaining 30% as the validation set. The training set is used for model parameter learning and updating, while the validation set is used to monitor model performance during training, guide training termination, and evaluate the final prediction accuracy of the model after training.

[0056] For the training and validation sets, corresponding training and validation samples were constructed in parallel, strictly following the method described in step S2 of Example 1. Each sample contains a fused feature vector generated from historical time-series window data, and a corresponding label for the actual hot water consumption in the future period.

[0057] S82. Iteratively train the CNN-LSTM prediction model using training samples. During training, the fused feature vector generated by the aforementioned method is used as the model input, and the corresponding actual hot water consumption is used as the target. The model uses mean squared error (MSE) or mean absolute percentage error (MAPE) as the loss function, and the Adam optimizer is used to iteratively train the CNN-LSTM prediction model. In each iteration, the model calculates the predicted value and loss through forward propagation, and updates the network weights through the backpropagation algorithm.

[0058] S83. During iterative training, the performance of the current model is periodically evaluated using validation samples, and its loss value on the validation set is calculated. The system continuously monitors the changes in this validation loss. When the validation loss value no longer decreases significantly over several consecutive training epochs, i.e., convergence is achieved, the training process is paused. The system saves the model parameters that perform best on the validation set at this point as the current optimal model.

[0059] S84. Using Mean Absolute Percentage Error (MAPE) as the core evaluation metric, the system calculates the prediction accuracy of the current best model on the entire validation set. The system then compares the calculated MAPE value with a pre-set performance threshold (e.g., 85% accuracy corresponds to approximately 15% MAPE): If the prediction accuracy is not lower than the preset threshold, the model is considered to have been successfully trained, and the current best model is determined as the final deployable model, thus completing the training.

[0060] If the prediction accuracy is lower than the preset threshold, it indicates that the model performance does not meet the requirements. At this time, the system will initiate an optimization process, return and optimize the feature construction strategy (e.g., adjust the feature encoding method, introduce new feature terms, etc.). After the strategy optimization, the system will re-execute the model training steps (i.e., starting from S82) using the original training and validation set data based on the updated feature construction method until a model that meets the accuracy requirements is trained.

[0061] Example 3

[0062] This embodiment provides a dormitory hot water consumption prediction system that integrates student schedules and weather conditions. The system is an integrated hardware and software platform that, through the collaborative work of its modules, automatically executes the entire process from data acquisition to prediction scheduling and self-optimization.

[0063] This system mainly consists of the following five functional modules, which are connected through standardized data interfaces: (1) Data Acquisition and Preprocessing Module: This module is responsible for acquiring raw data from multiple internal and external data sources in real time. It includes data interfaces with dormitory smart water meters and access control systems to collect real-time hot water consumption time-series data and student return-to-dormitory records; API connections with the campus academic affairs system and academic calendar system to obtain timetable data, holiday and exam week information; and a meteorological data interface for periodically acquiring weather forecast data. This module has a built-in data cleaning subroutine that performs timestamp alignment, missing value filling, and outlier filtering on the collected multi-source, heterogeneous data, outputting standardized, aligned, high-quality data packets to downstream modules.

[0064] (2) Feature Engineering Module: This module is responsible for converting raw data into numerical features that the model can understand. It receives standardized data packets from upstream and incorporates all the processing logic described in step S2 of Example 1. This module contains three parallel processing units: one for constructing a sleep intensity feature vector (calculating estimated sleep rate, determining water-sensitive periods, and generating date identifiers); one for standardizing weather data to construct a weather feature vector; and one for extracting periodic and trend features from historical consumption data to construct a hot water consumption time-series feature vector. Finally, a fusion unit concatenates the three vectors to generate a unified fused feature vector, which is then output to the prediction model module.

[0065] (3) Prediction Model Module: Responsible for performing the core prediction calculations. The core of this module is a pre-trained CNN-LSTM prediction model embedded in the software. This model includes a CNN sub-module with depthwise separable convolutions, a temporal attention sub-module, and an LSTM sub-module. This module receives fused feature vectors and automatically completes a series of calculations such as temporal sequence construction, feature extraction, attention weighting, and temporal modeling, ultimately outputting a scalar value, namely the predicted value of hot water consumption for a specified future period.

[0066] (4) Result Output and Scheduling Suggestion Module: This module is responsible for translating the prediction results into actual actions. It receives the predicted values ​​from the prediction model module. The module has a pre-built scheduling rule base and conversion algorithm that matches the parameters of specific hot water supply equipment (such as boilers, pumps, and valves). Based on the predicted values, this module can automatically generate optimal equipment control instructions (such as set temperature, start / stop time, and valve opening) or detailed management suggestion text, and push these instructions in real-time and accurately to the programmable logic controller of the dormitory hot water supply system or the centralized management platform of campus logistics through industrial communication protocols or message queues, driving the physical system to execute.

[0067] (5) Model Optimization Module: This is a service module that runs continuously in the background to ensure the system maintains high accuracy over a long period. This module continuously collects actual hot water consumption data from smart water meters and compares it with historical predictions to calculate errors. When the error accumulates to a certain scale or reaches a preset period, this module evaluates the model's current performance. If the performance is substandard, this module triggers an optimization process: either guiding the feature engineering module to adjust its feature construction strategy, or initiating an offline training process to generate a new model file. Finally, this module will safely and seamlessly update the verified and effective optimization strategy or the new model file to the production environment, replacing the original components.

[0068] During runtime, the above modules do not simply operate sequentially, but rather form a highly efficient collaborative network: Main prediction flow: Data acquisition → Feature engineering → Model prediction → Result output. This flow runs automatically and periodically at a high frequency (e.g., every 5 minutes) to achieve real-time prediction and scheduling.

[0069] Background optimization flow: The model optimization module runs independently and continuously, monitors system performance, and triggers training and updates when necessary, forming a slow but crucial adaptive closed loop.

[0070] Numerous specific details are set forth in the specification provided herein. However, it will be understood that embodiments of the invention may be practiced without these specific details. In some instances, well-known methods, structures, and techniques have not been shown in detail so as not to obscure the understanding of this specification.

[0071] Furthermore, those skilled in the art will understand that although some embodiments herein include certain features included in other embodiments but not others, combinations of features from different embodiments are intended to be within the scope of the invention and form different embodiments. Any of the claimed embodiments can be used in any combination.

Claims

1. A method for predicting dormitory hot water consumption by integrating student schedules and weather, characterized in that, Includes the following steps: S1. Obtain the data required for prediction, including: hot water consumption time series data of the target dormitory within a preset time window before the prediction time, as well as student daily routine data and weather data used to characterize the state of the future prediction period; S2. Extract and quantify features from the student daily routine data to construct a daily routine intensity feature vector; preprocess the weather data to construct a weather feature vector; preprocess the hot water consumption time series data to extract its hot water consumption time series features; and fuse the hot water consumption time series features, daily routine intensity feature vector, and weather feature vector to generate a fused feature vector. S3. Input the fused feature vector into the trained CNN-LSTM prediction model and output the predicted value of hot water consumption for a specified future period.

2. The method for predicting dormitory hot water consumption based on student schedules and weather as described in claim 1, characterized in that, In step S1, the student schedule association data includes time attribute data and behavior association data; the time attribute data includes at least the information on holidays, weekends and exam weeks as stipulated in the school calendar; the behavior association data includes at least the dormitory access control system's return-to-dormitory records and the timetable data corresponding to the future predicted period; the weather data includes temperature, humidity, wind speed and precipitation data.

3. The method for predicting dormitory hot water consumption based on student schedules and weather as described in claim 1, characterized in that, In step S2, the construction of the activity intensity feature vector includes: calculating the estimated in-room rate corresponding to the future prediction period based on the dormitory return record; determining whether the future prediction period is in the concentrated water use sensitive period after the end of the course based on the time schedule data, and generating corresponding identification features; and generating time identification features representing holidays, weekends and exam weeks based on the time attribute data. The estimated sleep rate, the water-sensitive period marker, and the time marker are concatenated, and the concatenated feature vector is standardized to form the sleep intensity feature vector.

4. The method for predicting dormitory hot water consumption based on student schedules and weather as described in claim 1, characterized in that, In step S2, the extraction of hot water consumption time-series characteristics includes extracting its periodic characteristics and trend characteristics; The periodicity features include at least daily and weekly fluctuation pattern encoding.

5. The method for predicting dormitory hot water consumption by integrating student schedules and weather as described in claim 1, characterized in that, In step S3, the CNN model of the CNN-LSTM prediction model adopts a depthwise separable convolutional structure to efficiently extract local patterns related to water use behavior from the fused feature vector.

6. The method for predicting dormitory hot water consumption based on student schedules and weather as described in claim 5, characterized in that, The CNN-LSTM prediction model also includes a temporal attention module; the temporal attention module is used to assign dynamic weights to the input temporal feature sequence to enhance the model's attention to peak water usage periods, and the temporal feature sequence is composed of the fused feature vectors of multiple time steps in chronological order.

7. The method for predicting dormitory hot water consumption based on student schedules and weather as described in claim 6, characterized in that, The workflow of the temporal attention module includes: S71. Receive the time-series feature sequence as input; S72. Calculate the attention score for each time step in the temporal feature sequence; S73. Based on the attention scores, the features at each time step are weighted and summed to generate a context vector; S74. Fuse the context vector with the temporal feature sequence to output a weighted temporal feature sequence.

8. The method for predicting dormitory hot water consumption based on student schedules and weather, according to any one of claims 1-7, is characterized in that... It also includes the model training step: S81. Collect time-series data on hot water consumption, student daily routine data, and weather data of the target dormitory during historical time periods, and divide them into training set and validation set, and construct training samples and validation samples according to the method described in claims 1 to 7. S82. Using the training samples, with mean squared error or mean absolute percentage error as the loss function, use the Adam optimizer to iteratively train the CNN-LSTM prediction model; S83. During iterative training, monitor model performance based on the validation samples, and save the current optimal model when the model loss function converges; S84. Using the mean absolute percentage error as the evaluation index, calculate the prediction accuracy of the current best model on the validation set; if the prediction accuracy is not lower than a preset threshold, then complete the training and determine the final model; otherwise, after optimizing the feature construction strategy, repeat step S82.

9. The method for predicting dormitory hot water consumption based on student schedules and weather, according to any one of claims 1-8, is characterized in that... Also includes: S4. Application steps: Convert the predicted hot water consumption for the specified future period into control parameters or scheduling suggestions for the dormitory hot water supply system, and output them to the logistics management platform; S5. Feedback optimization step: Collect actual hot water consumption data, compare it with the predicted value, and use the difference data to update the training dataset of the CNN-LSTM prediction model to achieve iterative optimization of the model.

10. A dormitory hot water consumption prediction system integrating student schedules and weather, characterized in that, For implementing the dormitory hot water consumption prediction method that integrates student schedules and weather as described in any one of claims 1-9, the system comprises: The data acquisition and preprocessing module is used to acquire and process hot water consumption time-series data, student daily routine data, and weather data; The feature engineering module is used to construct and fuse the activity intensity feature vector, weather feature vector, and hot water consumption time series features to generate a fused feature vector. The prediction model module has a built-in CNN-LSTM prediction model trained by the steps described in claim 8, which is used to receive the fused feature vector and output the prediction value. The results output and scheduling suggestion module is used to convert the predicted values ​​into scheduling instructions for hot water supply equipment. The model optimization module is used to collect actual hot water consumption data, compare it with the predicted value, and trigger strategy optimization of the feature engineering module or retraining of the prediction model module based on the comparison result.