Construction hoist dynamic energy-saving scheduling method based on AI load-working condition prediction

Abstract

The application discloses a construction hoist dynamic energy-saving scheduling method based on AI load-working condition prediction, which is used for controlling the operation of a construction hoist and comprises the following steps: data acquisition and feature construction, AI load-working condition prediction, multi-target optimization decision, execution control and closed-loop updating; real-time operation data and historical working condition data of the construction hoist are collected, a load-working condition correlation model is constructed, and the load state and transportation demand in the next period are predicted in advance; a multi-target optimization model of "lowest energy consumption, optimal efficiency and minimum impact" is constructed, and the optimal speed curve and start-stop timing are generated in combination with parameters such as time period, load level and floor height; through a closed-loop self-learning mechanism, the model parameters are dynamically corrected, the construction site working condition changes are adapted, and finally, energy-saving, efficient and stable driverless scheduling operation is realized.

Description

Technical Field

[0001] This invention relates to the field of construction hoist control technology, specifically to a dynamic energy-saving scheduling method for construction hoists based on AI load-condition prediction. It is applicable to double-cage or single-cage intelligent construction hoists, especially Zhejiang Construction Machinery SC200 / 200SZN46 and similar variable frequency speed control intelligent construction hoists, achieving energy-saving, efficient and stable operation in driverless mode. Background Technology

[0002] Construction hoists, as core equipment for vertical transportation in building construction, account for a significant proportion of the total energy consumption on construction sites. The scheduling and control methods of traditional construction hoists and existing intelligent models have obvious limitations. First, traditional models use a fixed-speed operating mode, such as a constant 45m / min. Regardless of the cage load, floor level, or call frequency, the same speed is maintained, resulting in significant energy waste during high-speed operation when unloaded or lightly loaded, and increased equipment wear and tear during heavy-load operation due to the resulting impact. Second, some improved models use a simple load-based speed regulation mode, dividing the speed into 2-3 fixed levels based solely on the real-time cage load, without considering key parameters such as floor height, call density, and time-of-day conditions. This simplistic speed regulation strategy has limited energy-saving effects, typically not exceeding 20%. Furthermore, the scheduling of existing intelligent construction hoists often prioritizes "proximity priority" and "forward priority," focusing solely on transportation efficiency without prioritizing optimal energy consumption. This leads to high empty-running rates (approximately 25%) and frequent starts and stops, further increasing energy losses. Furthermore, existing technologies passively respond to floor call signals, failing to predict load status, transportation demand, and operational changes in the next cycle. This prevents proactive adjustments to operating strategies, resulting in both energy waste and low efficiency. Finally, existing scheduling strategies often use fixed parameters, unable to adapt to dynamic changes in construction progress, personnel flow, and material transportation. Long-term operation leads to energy savings and poor equipment adaptability. A search reveals that existing patents (such as CN104860149B and CN106044422) and industry solutions do not address these core shortcomings. For example, they fail to integrate AI load prediction with operational analysis, lack a multi-objective optimization model balancing "lowest energy consumption + optimal efficiency + minimal impact," do not dynamically adjust speed control strategies based on multiple parameters, and lack a closed-loop self-learning mechanism. Therefore, a predictive, highly efficient, and adaptive dynamic scheduling control method for construction hoists is urgently needed. Summary of the Invention

[0003] To address the above technical problems, this invention provides a dynamic energy-saving scheduling method for construction hoists based on AI load-condition prediction, which can predict in advance, optimize multiple objectives, and dynamically adapt. This method solves the problems of passive scheduling, poor energy-saving effect, inability to balance efficiency and energy consumption, and weak adaptability in traditional scheduling.

[0004] The present invention adopts the following technical solution: A dynamic energy-saving scheduling method for construction hoists based on AI load-condition prediction is used to control the operation of construction hoists. Its features include the following steps: Data Acquisition and Feature Construction: The operation data of the construction hoist is collected in real time, including at least the cage load data, current floor height, direction of travel, call signals on each floor, and current time period information; The operational data is synchronized and structured to construct a temporal feature vector that reflects the spatiotemporal-state-business scenario of the construction hoist, which serves as the input to the load-operating condition association model. AI Load-Condition Prediction: The constructed time-series feature vector is input into the load-condition correlation model; Using the load-condition correlation model, the static business logic coupling relationship and dynamic time-series dependency relationship between different parameters in the time-series feature vector are analyzed and learned to predict the load status and transportation demand of the next cycle in advance. Multi-objective optimization decision-making: With the comprehensive optimization objectives of minimizing energy consumption, optimizing transportation efficiency, and minimizing operational impact, the predicted load status and transportation demand are combined with the current time period and floor height as inputs. Under a preset constraint, the optimal operating speed curve is generated and the start and stop times are determined. Execution control: The optimal operating speed curve is sent to the frequency conversion control system of the construction hoist, and the operation of the cage is controlled according to the start and stop timing; Closed-loop update: The predicted load status is compared with the actual operating load in real time, and the prediction error is calculated; when the prediction error exceeds a preset threshold, the load-operating condition correlation model is dynamically corrected, and / or the optimization weights in the multi-objective optimization decision-making step are adjusted. By adopting the above technical solution, this invention constructs a complete closed loop of "prediction-optimization-execution-correction". First, by constructing high-dimensional time-series feature vectors from multi-source heterogeneous data and utilizing deep learning with an AI model, it breaks through the limitations of traditional passive response, enabling advance prediction of future transportation demand and providing decision-making lead time for dynamic energy-saving scheduling. Second, the multi-objective optimization model comprehensively considers energy consumption, efficiency, and impact, generating a continuously variable speed curve that balances multiple performance indicators, rather than fixed-gear speed regulation, shifting from a single efficiency-first approach to multi-objective collaborative optimization. Furthermore, the closed-loop update mechanism corrects the AI ​​model and optimization weights through real-time feedback, enabling the system to adaptively adapt to dynamically changing site conditions (such as from main construction to decoration stages) without manual intervention, ensuring long-term operational stability and energy-saving effects.

[0005] Preferably, in the data acquisition and feature construction steps, the construction method of the time-series feature vector further includes: aligning the operating data according to the acquisition timestamp using a unified time benchmark; constructing a multi-dimensional time-series matrix from the continuously acquired operating data according to a set time window and acquisition frequency; and when performing structured processing on the time-series feature vector, at least the following are included: constructing current load weight and load change rate features for the hoist cage load data; constructing classification feature channels for the operating direction and call signals on each floor using one-hot encoding; and constructing time period features for the current time period information using cyclic feature encoding. By adopting the above technical solution, a feature that can completely reflect the construction hoist in four dimensions—time, space, operating status, and business needs—is constructed by structuring the original data. The load change rate feature reflects the dynamic trend of the load, one-hot encoding enables discrete information such as direction and call signals to be effectively understood by AI, and cyclic feature encoding cleverly solves the problem of time periodicity (such as the proximity of 23:59 and 00:01). These multidimensional, structured, and high-information-density feature vectors provide a high-quality data foundation for subsequent LSTM models to perform deep learning and capture the complex relationships between parameters, which is key to improving prediction accuracy.

[0006] Preferably, when performing one-hot encoding on the elevator call signals for each floor, an independent elevator call status channel is constructed for each service floor, and an overall elevator call statistic is generated as a macro-level demand characteristic. The overall elevator call statistic includes the total number of calls, the number of calls going up, the number of calls going down, and the waiting time on each floor. By adopting the above technical solution, and by constructing an independent elevator call status channel for each floor, the model can accurately perceive the specific elevator call demand on each floor, achieving refined management of elevator call signals. Simultaneously, the introduction of macro-level statistics such as the total number of calls, directional distribution, and waiting time provides the model with a quantitative perception of the overall transportation load. This allows the model to understand both micro-level "point-to-point" requests and grasp the macro-level transportation demand heatmap, thereby more accurately predicting the operating direction, load distribution, and transportation peak in the next cycle, significantly improving the accuracy and comprehensiveness of the prediction.

[0007] Preferably, the load-condition correlation model is a lightweight LSTM neural network model. The method further includes: utilizing the historical operating condition dataset of the construction hoist, and through the processing method of the data acquisition and feature construction steps, constructing historical time-series feature vector samples for training the lightweight LSTM neural network model; through the internal memory units and gating mechanisms of the lightweight LSTM neural network model, learning and capturing the correlation between one or more of the following parameters in the time-series feature vector: the dynamic correlation between the changing trend of the cage load in the past and the future load state; the covariance between time-period characteristics and call demand density and load level; and the logical correspondence between the spatial distribution of call signals and the direction of operation. By adopting the above technical solution, the lightweight LSTM model, with its unique memory units and gating mechanisms, is inherently adept at processing time-series data and is very suitable for learning the dynamic changing patterns in the operating data of construction hoists. By training with historical data, the model can automatically learn and master a variety of complex patterns with sequential dependencies. For example, it can infer that the load will enter a "full load" state in the future based on the rapid upward trend of the load in the past few tens of seconds; it can predict that the transportation demand will increase significantly based on the surge in elevator call signals during a specific period (such as after 7:00 am during the morning rush hour); and it can predict that the operating direction of the next cycle will be upward based on the concentrated occurrence of upward elevator call signals on a certain floor.

[0008] Preferably, in the AI ​​load-condition prediction step, the predicted load state includes at least light load, heavy load, and full load states, and the transportation demand includes at least elevator call frequency and target floor distribution. The load-condition correlation model outputs the prediction result 5-10 seconds in advance based on the time-series feature vector. By adopting the above technical solution, the 5-10 second prediction window is crucial. This time span is sufficient for the subsequent multi-objective optimization model to perform complex calculations and generate optimal control commands, while also being short enough to ensure that the prediction result is highly consistent with the upcoming actual operating conditions. By classifying the load state (light / heavy / full load) and predicting elevator call frequency and target floors, the optimization model can obtain sufficient input information to formulate refined speed strategies. For example, when predicting that full load is imminent and high-frequency elevator call response is required, the speed can be adjusted in advance to balance efficiency and impact, thereby realizing a paradigm shift from "post-event response" to "pre-event planning" in scheduling.

[0009] Preferably, the multi-objective optimization decision-making step further includes dynamically adjusting the operating strategy based on the current time period: during peak hours, priority is given to ensuring transportation efficiency while also considering energy conservation; during off-peak hours, energy consumption and transportation efficiency are balanced; during low-peak hours, energy conservation is prioritized, and the hoist cage is controlled to run at low speed or enter a sleep state that can be quickly awakened when unloaded. By adopting the above technical solution, this method introduces time-period awareness, enabling the scheduling strategy to adaptively match the characteristics of pedestrian and material flow at different times of the construction site. During peak hours, worker passage efficiency is the primary concern, and the strategy tends to operate at high speed; during low-peak hours, there is almost no transportation demand, and the strategy enters a low-power mode, prioritizing energy conservation; during off-peak hours, the optimal balance is sought between the two.

[0010] Preferably, the multi-objective optimization decision-making step further includes: determining a base speed value based on the predicted load state; dynamically adjusting the upper limit of the base speed value based on the current floor height, wherein when the current floor height or the target floor height exceeds a preset height threshold, the upper limit of the base speed value is reduced; and finally generating a stepless speed regulation curve ranging from 0 to the equipment's rated maximum speed. By adopting the above technical solution, this solution achieves precise speed control. The load state determines the base speed, while the floor height is used to correct the upper speed limit. For example, when operating at high altitudes, reducing the upper speed limit can effectively reduce energy consumption and noise caused by high-altitude wind resistance and improve operational safety. The final generation of a stepless speed regulation curve from 0 to 46 m / min means that the hoist cage's operating speed is continuously variable, capable of smooth adjustment according to real-time needs, avoiding the impact and energy waste caused by fixed-gear shifting, and achieving truly intelligent and stable operation.

[0011] Preferably, in the closed-loop update step, the formula for calculating the prediction error δ is: In the formula: Load prediction error (%) This represents the actual load value (t). The AI ​​model predicts a load value (t), with a preset threshold of 5%. In the closed-loop update step, when the prediction error exceeds the preset threshold, the dynamic correction includes: constructing new time-series feature vector samples using the latest collected operational data, and incrementally training the load-condition correlation model to update its network weight parameters. By adopting the above technical solution, this step endows the method with "self-learning" and "adaptive" capabilities. By introducing a specific error calculation formula and a 5% threshold, quantitative monitoring of the AI ​​model's accuracy is achieved. When the prediction deviates from reality, the system automatically triggers incremental training, using the latest operational data (e.g., new operating conditions caused by changes in worker work habits or material stacking locations) to fine-tune the model, enabling the AI ​​model to continuously track and adapt to real-time dynamic changes at the construction site. This effectively avoids performance degradation caused by operating condition migration, ensuring energy-saving effects and scheduling efficiency during long-term operation.

[0012] Preferably, in the multi-objective optimization decision-making step, the comprehensive optimization objective is expressed by a mathematical function, which is: Where F is the comprehensive optimization objective value, , , These are the weighting coefficients for energy consumption, efficiency, and impact, respectively; E is the energy consumption per unit time; v is the operating speed; and a is the start-stop impact acceleration. Furthermore, the "adjusting the optimization weights in the multi-objective optimization decision-making step" in the closed-loop update step includes dynamic adjustment. , , The value of . By adopting the above technical solution: the present invention concretizes the abstract target into a computable mathematical model, through weight coefficients ( , , This achieves a quantification of the importance of the three objectives. More importantly, the closed-loop update mechanism dynamically adjusts these weights; for example, it automatically increases the weight when excessive energy consumption is detected. This makes the optimizer more focused on energy saving; when it detects a slow response, it automatically increases the power consumption. This improves operating speed. This adaptive weight adjustment mechanism enables the scheduling strategy to dynamically and intelligently find the optimal balance point among the three conflicting objectives under different operating conditions, achieving both flexibility and optimality of the strategy.

[0013] Preferably, the system also includes a command safety verification and arbitration step, which is set between the multi-objective optimization decision-making step and the execution control step. A safety arbitration module acquires and verifies whether the generated optimal operating speed curve and start / stop timing meet preset safety conditions. If they do, the parameters are sent to the frequency converter control system; otherwise, they are rejected or corrected to ensure safety priority. The system also includes an extreme condition correction step. When a critical sensor failure or data loss is detected, the following strategies are executed: immediately control the hoist cage to decelerate, stop, and dock at the nearest floor; trigger a fault alarm and automatically switch to a preset emergency operation mode to meet basic safety requirements; after the fault is recovered, automatically verify the data accuracy; if the accuracy meets the standard, automatically switch back to the dynamic energy-saving scheduling method; if the accuracy does not meet the standard, maintain the emergency mode. By adopting the above technical solution, safety is the lifeline of industrial control. This method, by introducing the "command safety verification and arbitration" step, sets up a safety "firewall" for scheduling commands while the optimization algorithm pursues ultimate energy saving and efficiency, ensuring that any control command is within the safety boundary, effectively resolving potential conflicts between energy-saving optimization and safety assurance. Meanwhile, an "extreme condition correction" mechanism was designed to address potential extreme conditions such as sensor failures and communication interruptions. The system can automatically degrade to emergency safety mode and seamlessly switch back to intelligent mode after normal operation resumes. This multi-layered safety assurance system of "prioritizing safety, intelligent degrading, and automatic recovery" greatly enhances the system's robustness and reliability, enabling it to adapt to complex on-site environments. This is a key guarantee that this method can move from the laboratory to practical application.

[0014] Compared with the prior art, the present invention has the following advantages: (1) Significant energy saving effect: The overall energy saving rate reaches 22%-28%. Compared with the traditional fixed speed operation mode, a single construction hoist saves about 18,000 kWh of electricity per year, reduces carbon emissions by about 13.5 tons, and lowers construction energy consumption costs; (2) Improve operational efficiency: Through AI prediction and optimized scheduling, the empty running rate is reduced to ≤15%, the number of start-stop times is reduced by about 30%, and the elevator call response time is shortened by about 15%, meeting the high-intensity transportation needs of construction sites; (3) Extend equipment life: Optimize speed curve and start-stop timing, reduce operating shock, reduce wear and tear on components such as motors, brakes, and guide rails, extend equipment life by about 20%, and reduce equipment maintenance costs; (4) Strong adaptability: Through the closed-loop self-learning mechanism, it can adapt to the working conditions of different construction sites and different construction stages (from main construction to decoration stage), without the need for manual parameter adjustment. It is compatible with Zhejiang Construction Machinery SC200 / 200SZN46 and similar intelligent construction hoists, without the need for additional hardware modification. (5) Achieve driverless adaptation: It is perfectly integrated with the driverless operation mode, eliminating the need for a dedicated driver, further reducing labor costs, and avoiding problems such as inaccurate speed control and energy waste caused by manual operation. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating the overall process of the method of the present invention. Detailed Implementation

[0016] To facilitate understanding of the technical solution of the present invention, the following detailed description is provided in conjunction with the accompanying drawings and specific embodiments.

[0017] Example 1

[0018] One specific embodiment of this invention is based on the Zhejiang Construction Machinery SC200 / 200SZN46 intelligent construction hoist, but is not limited thereto. This method is also applicable to similar variable frequency speed-regulating intelligent construction hoists from other brands. In this embodiment, the construction hoist has a double-cage structure, a rated load of 2t / 2t, and a variable frequency adjustable lifting speed of 0-46m / min, supporting driverless automatic operation. Its hardware configuration includes: a weight sensor (accuracy ±1%) for collecting real-time load data of the hoist cage; an encoder (accuracy ±1mm) and laser positioning module for collecting current floor height and running direction data; wireless external call panels for collecting call signals on each floor; a system clock module for acquiring time period information; a data storage module for storing historical data; an AI control unit (equipped with a lightweight LSTM neural network and a multi-objective optimization model) as the core computing unit; and a variable frequency control system for executing final instructions. All modules communicate via industrial Ethernet to ensure real-time data transmission and reliable transmission of control commands.

[0019] like Figure 1 As shown, the dynamic energy-saving scheduling method for construction hoists based on AI load-condition prediction of the present invention is implemented through the following steps: Step S100: Data Acquisition and Feature Construction. The AI ​​control unit collects data from various sensors in real time, including cage load data, current floor height, direction of travel, call signals from each floor, and current time period information. To transform this multi-source heterogeneous data into structured input that the AI ​​model can learn efficiently, this method performs time-series synchronization and feature construction processing on the data. Using a unified time reference (such as the system clock), various types of data are aligned according to the acquisition timestamp (10Hz acquisition frequency). A time window is set (e.g., 30 seconds), and all aligned data within this window are constructed into a multi-dimensional time-series matrix. In the matrix, for load data, not only is the current load weight recorded, but a load change rate feature is also constructed through differential calculation. For the direction of operation, one-hot encoding is used to form three feature channels: [upward (0 / 1), downward (0 / 1), stopped (0 / 1)]. For call signals on each floor, an independent call status channel is constructed for each service floor (e.g., 1 represents a call, 0 represents no call), and the total number of calls, the number of upward / downward calls, and the waiting time on each floor are calculated simultaneously as macroscopic statistical features. For the current time period information, sine and cosine cyclic feature encoding is used to smoothly represent the periodicity of time. Finally, a temporal feature vector T that can simultaneously reflect the spatial location, operating status, business needs, and temporal background of the construction hoist is constructed as the input of the subsequent model. In this structure, by transforming the original, discrete, and heterogeneous operating data into a high-dimensional, structured temporal feature vector T, the subsequent AI model can learn in a unified environment, which is a prerequisite for achieving high-precision prediction. Meanwhile, by introducing processing methods such as load change rate, one-hot encoding, and cyclic feature encoding, not only is the original information of the data preserved, but also deep information such as dynamic trends, discrete classifications, and periodic patterns between data are extracted, which greatly enriches the input features, thereby improving the model's ability to characterize complex working conditions and bringing about a significant improvement in prediction performance.

[0020] Step S200: AI Load-Condition Prediction. A lightweight LSTM neural network is deployed in the AI ​​control unit as a load-condition correlation model M. First, using a historical operating condition dataset (at least 6 months of data), historical time-series feature vector samples are constructed in the same way as in Step S100. Model M is trained offline, setting hyperparameters such as learning rate, number of iterations, and number of hidden layer nodes to reduce the model's prediction error to within 5%. After training, model M is put into online prediction. It receives the real-time time-series feature vector T constructed in Step S100 in real time, and automatically learns and captures the complex correlations between different parameters in vector T using its internal memory units and gating mechanisms. For example, model M can learn: the dynamic correlation of the load rapidly increasing from 0.2t to 1.8t in the past few seconds, inferring that the load status in the next cycle is "full load"; it can learn the covariant pattern of the increase in elevator call signal density after 7:00 AM, predicting a surge in transportation demand; and it can learn the logical correspondence that when multiple floors have upward elevator call signals, the next cycle's operating direction is very likely to be "upward". Based on this, model M outputs predictions of the load status (light load, heavy load, full load) and transportation demand (elevator call frequency, target floor distribution) for the next cycle 5-10 seconds in advance. In this structure, by utilizing the temporal memory capability of the LSTM neural network, model M no longer passively processes the data of the current moment, but actively "memorizes" and "understands" the evolution of the working conditions over a period of time, and makes future predictions based on this, which fundamentally changes the passivity of scheduling decisions. At the same time, by capturing deep correlations such as load trends, time period patterns, and spatial distribution of elevator calls, the prediction results are more intelligent and accurate, and can accurately predict non-periodic but predictable working conditions such as "workers are about to leave work, and there will be a large number of downward elevator calls on each floor" or "materials have been stacked and a large-scale upward transportation is about to occur," significantly improving the foresight of the scheduling strategy.

[0021] Step S300: Multi-objective optimization decision. A multi-objective optimization model O is constructed internally within the AI ​​control unit, with the objective function being: Where F is the comprehensive optimization target value, ω1, ω2, and ω3 are the weighting coefficients of energy consumption, efficiency, and impact, respectively (ω1+ω2+ω3=1), E is the energy consumption per unit time (kWh), v is the operating speed (m / min), and a is the start-stop impact acceleration (m / s²). 2The constraints of the optimization model include: lifting speed 0-46 m / min, leveling accuracy ±2 mm, and start / stop impact ≤0.5 g. Model O receives the prediction results (load status, transportation demand) from step S200, and combines them with the current time period (peak / off-peak / valley) and the current floor height. It then uses intelligent algorithms such as particle swarm optimization to solve the problem, generating an optimal operating speed curve with stepless speed regulation from 0 to 46 m / min and the start / stop timing of each action. For example, at peak time 7:30 AM, with a predicted light load and a target floor of the 18th floor (54m high), the optimization algorithm generates a curve that accelerates to 40m / min (the speed limit is automatically reduced due to the height > 50m) and then decelerates to a stop. At off-peak time 2:00 PM, with a predicted heavy load and downward movement, a curve is generated that descends smoothly at 30m / min to avoid high-speed impact from heavy load. At off-peak time 10:00 PM, with no load and no elevator calls, a curve is planned that moves to the first floor at 8m / min and enters a dormant state. In this structure, the multi-objective optimization model O quantifies the three conflicting objectives of energy consumption, efficiency, and impact through mathematical expressions. It can flexibly change decision preferences under different time periods and operating conditions by adjusting the weight coefficients ω1, ω2, and ω3, thereby generating a speed regulation curve that takes into account multiple performance indicators. This truly realizes the shift from a single efficiency-first scheduling paradigm to multi-objective collaborative optimization. This continuous stepless speed regulation strategy avoids the mechanical shock and secondary energy waste caused by traditional fixed-gear transmission. At the same time, it combines factors such as floor height to make "refined" corrections to the speed, making the operation of the hoist cage more stable, efficient and energy-saving. This brings unexpected technical effects of "triple" benefits: extended equipment life, improved ride comfort and further reduced energy consumption.

[0022] Step S400: Command Safety Verification and Arbitration. Before the optimal speed curve and start / stop timing generated in step S300 are sent to the actuator, a mandatory safety verification performed by the safety arbitration module A is required. Module A acquires and verifies whether the received scheduling command meets preset safety conditions. These preset safety conditions include at least: whether the current cage door area is completely closed, whether the safety doors on each floor are locked, whether there are emergency signals such as an emergency stop button being pressed, and whether any obstacles are detected in the operating corridor. If all safety conditions are met, the command passes the verification, and the process proceeds to step S500. If any condition is not met, the safety arbitration module A will reject or forcibly correct the command, for example, by limiting the target speed to a very low safety threshold (such as 5 m / min) or forcibly inserting a parking section and waiting for the safety conditions to be lifted. In this structure, by setting mandatory command safety verification and arbitration steps, a safety control between energy-saving scheduling and operational safety is constructed. Even if the intelligent optimization algorithm generates mathematically optimal instructions that may physically violate safety boundaries in pursuit of ultimate energy saving or efficiency, the safety arbitration module A can intercept or correct them at the last minute, thus prioritizing safety and eliminating potential safety risks caused by algorithm failures or boundary conditions. This step ensures that the entire dynamic energy-saving scheduling method operates within an absolutely safe framework while providing excellent performance.

[0023] Step S500: Execution Control. The AI ​​control unit converts the optimal operating speed curve, after safety verification, into frequency converter control commands and sends them to the frequency converter control system of the construction hoist. The frequency converter adjusts the output frequency in real time according to the commands, drives the motor, and controls the hoist cage to run precisely according to the planned speed curve and accurately position itself. At the same time, the system automatically responds to elevator call signals from each floor and, combined with the transportation demand predicted in step S200 (such as floor call density and direction), optimizes the elevator call scheduling sequence (such as prioritizing nearby and forward calls, or reserving capacity for upcoming heavy-load calls), achieving fully automatic energy-saving scheduling in driverless mode.

[0024] Step S600: Closed-loop update. The AI ​​control unit compares the predicted data from step S200 with the measured data fed back after step S500 in real time, and calculates the prediction error. The formula for calculating the prediction error is: ,in Load prediction error (%) This represents the actual load value (t). The AI ​​model predicts the load value (t), with a preset threshold of 5%. When δ exceeds the preset threshold of 5%, the system automatically performs closed-loop correction: First, it dynamically corrects the load-condition correlation model M by using the latest collected operational data from the past 7 days (for example, when it is found that the weight of materials to be transported has significantly decreased as the main construction stage transitions to the decoration stage), incrementally training the LSTM model and updating its network weight parameters so that its prediction accuracy returns to within 5%; Second, it dynamically adjusts the weight coefficients ω1, ω2, and ω3 of the multi-objective optimization model O. For example, if the energy consumption deviation rate ΔE is detected to be high, ω1 (energy consumption weight) is automatically increased; if the elevator call response delay rate ΔT is detected to be high, or the time period changes from off-peak to peak, ω2 (efficiency weight) is automatically increased, and vice versa. The adjustment process follows the principles of smooth update and normalization to ensure that the scheduling strategy always approaches the optimal balance point under the current operating conditions. In this structure, the closed-loop update mechanism enables the system to have the ability to learn and evolve on its own. By continuously monitoring the deviation between its predictions and execution, and proactively adjusting the AI ​​model and optimization objectives accordingly, this method can automatically adapt to any dynamically changing working conditions, such as construction progress, personnel flow, and material transportation. This completely solves the common problems of "performance degradation" and "poor adaptability" faced by traditional fixed-parameter models after long-term operation. More importantly, this mechanism can cope with gradual changes in working conditions (such as from main structure to decoration) and sudden mode changes without any manual intervention, greatly reducing operation and maintenance costs and improving the system's robustness and long-term operational reliability. This is another core advantage of this method compared to existing technologies.

[0025] To further refine the closed-loop update process, let's illustrate with specific numerical examples: During a certain run, the AI ​​model predicted a load W. p ᵣ e =1.0t (belongs to light load, ≤50% of rated load, the rated load of this model is 2t), the actual load W collected by the weight sensor during actual operation. ant=1.1t (still considered light load, ≤50% of rated load). According to the above prediction error formula, δ=|(1.1-1.0) / 1.1|×100%≈9.1%, which exceeds the preset 5% threshold. At this time, the AI ​​control unit automatically triggers the model correction mechanism, using the latest operating data from the past 7 days (including 1200 sets of valid data such as various load levels, floor height, and elevator call frequency) to incrementally train the LSTM neural network. The focus is on adjusting the weight parameters of the fully connected layer of the network (changing the original weight coefficient from 0.32 to 0.36). The number of training iterations is set to 200. After training, the system is put back into operation. The predicted load under subsequent similar working conditions is corrected to 1.08t, while the actual load remains 1.1t. The prediction error is calculated as δ=|(1.1-1.08) / 1.1|×100%≈1.8%, successfully falling back to within the 5% threshold, ensuring that the model prediction accuracy continues to meet the standard and adapts to changes in working conditions.

[0026] Step S700: Extreme Condition Correction. When a fault is detected in a critical sensor (such as a weight sensor, encoder, or laser positioning module) or data communication packet loss causes the acquisition rate to fall below the predetermined level, the system immediately triggers safety protection: First, it immediately controls the hoist cage to decelerate, stop, and automatically dock at the nearest level floor to prevent misjudgments and dangerous actions caused by inaccurate data; then, it issues an alarm through the voice broadcast module and automatically switches the control system to the preset emergency manual mode, allowing on-site workers to perform basic lifting operations via physical buttons to ensure safe evacuation; finally, after fault diagnosis is completed and data is restored, the system automatically performs data accuracy verification. If the load accuracy is restored to ±1% and the floor accuracy is restored to ±1mm, the system automatically and seamlessly switches back to the automatic energy-saving scheduling mode; if the accuracy is not up to standard, it continues to maintain the emergency mode until the fault is completely eliminated.

[0027] Referring to the accompanying drawings, the working principle of this invention is as follows: After the method is started, the collected multi-source heterogeneous operating data is first constructed into a time-series feature vector T containing deep information through S100, and input into the load-condition correlation model M deployed with a lightweight LSTM neural network (S200). Model M uses the complex correlation between parameters it has learned to predict the load status (light / heavy / full load) and transportation demand in real time for the next 5-10 seconds. Subsequently, the multi-objective optimization model O (S300) finds the optimal balance point between energy consumption, efficiency and impact based on the prediction result, combined with the current time period and floor height, and solves for an optimal operating speed curve and start-stop timing for stepless speed regulation. After the curve passes the safety verification of the safety arbitration module A (S400), it is sent to the frequency converter control system for execution (S500). Throughout the operation, the system continuously compares predicted and measured values ​​through a closed-loop update mechanism (S600). If a prediction error exceeds the limit, the AI ​​model and optimization weights are immediately dynamically corrected to ensure the system can adaptively keep up with changes in working conditions. Simultaneously, an extreme working condition correction mechanism (S700) serves as a final safety line, ready to respond to unforeseen circumstances such as sensor failures. Therefore, through a complete closed loop of "prediction-optimization-verification-execution-correction," this invention combines AI prediction with multi-objective optimization, supplemented by safety verification and self-learning mechanisms. This fundamentally solves the problems commonly found in traditional construction hoist scheduling, such as passive response, poor adaptability, and the inability to balance energy saving and efficiency. It achieves efficient, energy-saving, stable, safe, and intelligent automated operation in driverless mode, and is perfectly compatible with specific models of variable frequency speed-regulating intelligent construction hoists, resulting in significant economic and social benefits.

[0028] Example 2

[0029] The closed-loop update step includes... , , Adjustment of the value.

[0030] Specifically, it is a rule-based hierarchical fuzzy adjustment (intuitive and fast response); this method establishes a fuzzy rule base to adjust the weights based on the source of prediction error and operating conditions.

[0031] Step 1: Calculate the evaluation indicators In each closed-loop update cycle (e.g., after N runs or a fixed duration), calculate the following inputs: Load prediction error δ.

[0032] Energy consumption deviation rate ΔE: The deviation between actual energy consumption per unit time and theoretical optimal energy consumption.

[0033] Elevator call response delay rate ΔT: The deviation between the actual average elevator call response time and the target value.

[0034] Number of impact exceedances N_a: The number of times the start-stop impact acceleration exceeds the comfort threshold.

[0035] Step 2: Fuzzification and Rule-Based Reasoning The above δ, ΔE, ΔT, and N_a are fuzzed into {low, medium, high}, or a threshold is directly set, and then Δω is adjusted according to the rules: Rule 1 (Energy Consumption Priority): If ΔE is high (actually too power-consuming) and ΔT is low (efficiency margin is still available), then Δω1'+0.1 (increase the weight of the energy consumption target) and Δω3'-0.05 (slightly relax the impact constraint).

[0036] Rule 2 (Efficiency First): If ΔT is high (response is too slow, complaints have been received) or the time period switches to peak, then Δω2'+0.15 (increase the weight of speed / efficiency target), Δω1'-0.1.

[0037] Rule 3 (Smoothness Priority): If N_a is high (excessive impact) or δ fluctuates repeatedly (unstable operating conditions), then Δω3'+0.1 (increase the impact target weight) and Δω2'-0.05 (allow speed fluctuations in exchange for smoothness).

[0038] Step 3: Smooth Update and Normalization To avoid oscillations in the velocity curve caused by sudden changes in weights, smooth updates and forced normalization are adopted: Here, η is the learning rate (e.g., 0.3), ensuring a smooth transition of weights. Normalization ensures that the updated ∑ω... i =1.

[0039] Example 3

[0040] The closed-loop update step includes... , , Adjustment of the value.

[0041] Specifically, it is gradient-based adaptive optimization (precise and interpretable). Treating weight adjustment as a higher-level optimization problem, we minimize a "higher-level loss function" using gradient descent.

[0042] Design a loss function L that can comprehensively reflect the current state of dissatisfaction with the system: Where α, β, and γ are the importance coefficients (fixed) of each sub-objective. E act This is the current actual energy consumption, E target It is the expected energy consumption, T resp,actT represents the actual elevator call response time. resp,target N represents the target elevator call response time. a,act N represents the actual number of times the impact exceeded the limit. a,target The number of times the target system exceeds its limit. This loss function measures the weighted distance between the current system performance and the ideal performance.

[0043] Step 2: Calculate the gradient of the weights By sampling actual operating data, the impact of unit weight changes on the final operating result (E, T, a) is estimated, and then the gradient of the loss function L with respect to the weight ωi is calculated.

[0044] Step 3: Update the weights along the gradient direction Here, λ is the adjustment step size. Finally, normalization is performed to ensure that ∑ωi=1.

[0045] Example 4

[0046] The closed-loop update step includes... , , Adjustment of the value.

[0047] Specifically, it is a timed-driven strategy snapshot switching method (stable and reusable). This method links weight adjustments to the "operating conditions" of different time periods, and forms an optimal pattern library by accumulating experience.

[0048] Step 1: Create a "Policy Snapshot" database The system records the optimal set of data after manual or automatic optimization under different typical operating conditions (such as "morning peak - heavy load - high-rise" and "midday off-peak - light load - low-rise"). , , The values ​​are used to form a snapshot library of "operating conditions-optimal weights".

[0049] Step 2: Online matching and switching During the closed-loop update cycle, gradients are no longer calculated in real time. Instead, the current operating mode (predicted load + current time period + call density) is identified, the most matching mode is retrieved from the snapshot library, and its corresponding weight is directly called.

[0050] Step 3: Online Evolution of the Snapshot Library If, under a certain mode, the L(ω) loss function fails to decrease after a period of time, a small-scale local optimization (such as single-step gradient descent in method two) is triggered to find better weights under that mode and update the snapshot library. This method balances the stability of the policy with its self-evolutionary capability.

[0051] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention shall be determined by the scope defined in the claims. Any improvements and modifications made by those skilled in the art without departing from the spirit and scope of the present invention shall also be considered as within the scope of protection of the present invention.

Claims

1. A dynamic energy-saving scheduling method for construction hoists based on AI load-condition prediction, used to control the operation of construction hoists, characterized in that... Includes the following steps: Data Acquisition and Feature Construction: The operation data of the construction hoist is collected in real time, including at least the cage load data, current floor height, direction of travel, call signals on each floor, and current time period information; The operational data is synchronized and structured to construct a temporal feature vector that reflects the spatiotemporal-state-business scenario of the construction hoist, which serves as the input to the load-operating condition association model. AI Load-Condition Prediction: The constructed time-series feature vector is input into the load-condition correlation model; Using the load-condition correlation model, the static business logic coupling relationship and dynamic time-series dependency relationship between different parameters in the time-series feature vector are analyzed and learned to predict the load status and transportation demand of the next cycle in advance. Multi-objective optimization decision-making: The predicted load status and transportation demand are combined with the current time period and floor height as inputs to generate the optimal operating speed curve and determine the start and stop times under a preset constraint. Execution control: The optimal operating speed curve is sent to the frequency conversion control system of the construction hoist, and the operation of the cage is controlled according to the start and stop timing; Closed-loop update: The predicted load status is compared with the actual operating load in real time, and the prediction error is calculated; when the prediction error exceeds a preset threshold, the load-operating condition correlation model is dynamically corrected, and / or the optimization weights in the multi-objective optimization decision-making step are adjusted.

2. The dynamic energy-saving scheduling method for construction hoists based on AI load-condition prediction according to claim 1, wherein the construction method of the time-series feature vector in the data acquisition and feature construction step further includes: The operational data is aligned according to the collection timestamp using a unified time base; The continuously collected operational data is constructed into a multi-dimensional time series matrix according to the set time window and collection frequency; When performing structured processing on the time-series feature vectors, at least the following are included: For the load data of the hoist cage, construct the current load weight and load change rate characteristics; For the direction of travel and the call signals for each floor, a classification feature channel is constructed using one-hot encoding; For the current time period information, a time period feature is constructed using cyclic feature encoding.

3. The dynamic energy-saving scheduling method for construction hoists based on AI load-condition prediction according to claim 2, characterized in that, When performing one-hot encoding on the elevator call signals of each floor, an independent elevator call status channel is constructed for each service floor, and an overall elevator call statistic is generated as a macro demand feature; the overall elevator call statistic includes the total number of elevator calls, the number of upward elevator calls, the number of downward elevator calls, and the waiting time for each floor.

4. The dynamic energy-saving scheduling method for construction hoists based on AI load-condition prediction according to claim 1, characterized in that, The load-condition correlation model is a lightweight LSTM neural network model; The method further includes: using the historical operating condition dataset of the construction hoist, and through the processing method of the data acquisition and feature construction steps, constructing historical time-series feature vector samples for training the lightweight LSTM neural network model; Through the internal memory units and gating mechanism of the lightweight LSTM neural network model, the correlation between one or more of the following parameters in the temporal feature vector is learned and captured: The dynamic correlation between the past trend of cage load and future load status; Covariance between time-period characteristics and elevator call demand density and load level; The logical correspondence between the spatial distribution of elevator call signals and their direction of travel.

5. The dynamic energy-saving scheduling method for construction hoists based on AI load-condition prediction according to claim 1, characterized in that, In the AI ​​load-condition prediction step, the predicted load state includes at least light load, heavy load and full load states, and the transportation demand includes at least elevator call frequency and target floor distribution; the load-condition association model outputs the prediction result 5-10 seconds in advance based on the time-series feature vector.

6. The dynamic energy-saving scheduling method for construction hoists based on AI load-condition prediction according to claim 1, characterized in that, The multi-objective optimization decision-making step also includes dynamically adjusting the operating strategy based on the current time period: During peak hours, priority should be given to ensuring transportation efficiency while also taking energy conservation into account; During off-peak hours, balance energy consumption with transportation efficiency; During off-peak hours, energy conservation is prioritized, and the hoist cage is controlled to run at low speed or enter a hibernation state that can be quickly awakened when unloaded.

7. The dynamic energy-saving scheduling method for construction hoists based on AI load-condition prediction according to claim 6, characterized in that, The multi-objective optimization decision-making step also includes: The base speed value is determined based on the predicted load condition; The upper limit of the speed base value is dynamically adjusted according to the current floor height, wherein when the current floor height or the target floor height exceeds a preset height threshold, the upper limit of the speed base value is reduced. The final stepless speed regulation curve is generated within the range of 0 to the rated maximum speed of the equipment.

8. The dynamic energy-saving scheduling method for construction hoists based on AI load-condition prediction according to claim 1, characterized in that, In the closed-loop update step, the formula for calculating the prediction error δ is: In the formula: Load prediction error (%) The actual load value (t). The AI ​​model is used to predict the load value (t), and the preset threshold is 5%; In the closed-loop update step, when the prediction error exceeds the preset threshold, the dynamic correction includes: constructing new time-series feature vector samples using the latest collected operating data, and incrementally training the load-condition correlation model to update its network weight parameters.

9. The dynamic energy-saving scheduling method for construction hoists based on AI load-condition prediction according to claim 1, characterized in that, In the multi-objective optimization decision-making step, the comprehensive optimization objective is expressed by a mathematical function, which is: Where F is the comprehensive optimization objective value, , , These are the weighting coefficients for energy consumption, efficiency, and impact, respectively; E is the energy consumption per unit time; v is the operating speed; and a is the start-stop impact acceleration. Furthermore, the "adjusting the optimization weights in the multi-objective optimization decision-making step" in the closed-loop update step includes dynamic adjustment. , , The value of .

10. The dynamic energy-saving scheduling method for construction hoists based on AI load-condition prediction according to claim 1, characterized in that, It also includes a command safety verification and arbitration step, which is set between the multi-objective optimization decision step and the execution control step: a safety arbitration module obtains and verifies whether the generated optimal operating speed curve and start-stop timing meet the preset safety conditions. If they meet the conditions, they are sent to the frequency conversion control system. If they do not meet the conditions, they are rejected or corrected to ensure that safety takes priority. It also includes extreme condition correction steps, which execute the following strategies when a critical sensor failure or data loss is detected: Immediately slow down the hoist cage, bring it to a stop, and dock it at the nearest floor. Trigger a fault alarm and automatically switch to a preset emergency operation mode to meet basic safety requirements; After the fault is recovered, the data accuracy is automatically checked. If the accuracy meets the standard, the system automatically switches back to the dynamic energy-saving scheduling method; if the accuracy does not meet the standard, the emergency mode is maintained.

Images