Building electric power energy-saving control system based on load prediction and dynamic regulation and control

Abstract

The invention belongs to the technical field of intelligent building and energy management, particularly relates to a building electric power energy-saving control system based on load prediction and dynamic regulation and control, and aims to solve the problems of low energy utilization efficiency, lagging load response and difficulty in multi-target cooperative control of a building electric power system. The system comprises a load data acquisition module, a short-term load prediction module, a dynamic energy efficiency evaluation module, a multi-objective optimization scheduling module, and a real-time regulation and control execution and feedback correction module. High-precision load prediction is performed by fusing multi-source data, an energy-saving scheduling scheme is generated by combining energy efficiency evaluation and a hierarchical optimization algorithm, and closed-loop dynamic regulation and control are realized. Photovoltaic prediction, energy storage scheduling and a three-level time scale cooperation mechanism are preferably integrated, and the utilization rate of clean energy and the adaptive capacity of the system are improved. The comprehensive energy consumption can be reduced by more than 12%, the starting and stopping frequency of equipment is reduced, the service life of the equipment is prolonged, and the power supply reliability and the operation and maintenance intelligence level are improved.

Description

Technical Field

[0001] This invention belongs to the field of intelligent building and energy management technology, specifically relating to a building power energy-saving control system based on load forecasting and dynamic regulation. Background Technology

[0002] Building energy management systems, as a core component of modern intelligent buildings, are committed to achieving efficient energy utilization and sustainable development. Driven by accelerated urbanization, refined management of electricity consumption has become a crucial aspect of building operation and maintenance. Electricity load, as the primary form of energy consumption in building operation, presents significant challenges to grid stability and energy cost control due to its volatility and uncertainty. Therefore, how to achieve energy-saving optimization of building power systems through intelligent means has become an important research direction in the intersection of smart buildings and energy.

[0003] Among these technologies, data-driven load forecasting and real-time control are gradually becoming the mainstream approach for building power management. This technology aims to construct load change models by analyzing historical electricity consumption data, environmental parameters, and user behavior, and then dynamically adjust power supply strategies based on real-time monitoring information to achieve supply-demand balance and improved energy efficiency. Such systems typically integrate sensor networks, edge computing units, and a central control platform, supporting multi-level energy dispatch decisions and are widely used in large-scale building scenarios such as commercial buildings, hospitals, and industrial parks.

[0004] While existing technologies have achieved basic load data acquisition and simple threshold control, several bottlenecks remain: a lack of deep fusion capabilities for multi-source heterogeneous data (such as weather, work schedules, and equipment status) leads to insufficient load forecasting accuracy; control strategies often rely on fixed rules or delayed feedback, making them difficult to adapt to complex and ever-changing energy consumption scenarios; high system response latency prevents real-time dynamic matching of power supply and actual demand; and energy-saving control often comes at the cost of some comfort, lacking a multi-objective collaborative optimization mechanism for energy efficiency, economy, and user experience. These problems are particularly pronounced during peak electricity consumption periods or in scenarios with sudden load changes, easily leading to energy waste, equipment overload, or degraded power quality. Therefore, there is an urgent need for a building power energy-saving control system with high-precision forecasting capabilities and adaptive dynamic control characteristics. Summary of the Invention

[0005] The purpose of this invention is to provide a building power energy-saving control system based on load forecasting and dynamic regulation to solve the technical problems commonly found in existing building power systems, such as low energy utilization efficiency, lagging response to load fluctuations, poor equipment operation coordination, and lack of forward-looking regulation capabilities. Current building power systems mostly rely on fixed thresholds or empirical rules for start-up and shutdown control, which is difficult to adapt to complex and ever-changing electricity load demands. This results in power distribution equipment operating under suboptimal conditions for extended periods, causing significant energy waste and equipment damage. Especially in high-density electricity consumption scenarios such as commercial complexes and large public buildings, multiple types of loads, such as air conditioning, lighting, and elevators, operate simultaneously, exhibiting strong coupling and significant time-varying characteristics. Traditional control strategies cannot achieve refined and adaptive energy allocation and scheduling. Furthermore, existing systems have weak support capabilities for renewable energy access, coordination of energy storage devices, and grid interaction responses, hindering the intelligent evolution of building-level energy systems. Therefore, there is an urgent need to construct a new power energy-saving control system with accurate forecasting capabilities, a real-time dynamic regulation mechanism, and a multi-level collaborative optimization architecture.

[0006] The technical solution of this invention is a building power energy-saving control system based on load forecasting and dynamic regulation, comprising a load data acquisition module, a short-term load forecasting module, a dynamic energy efficiency assessment module, a multi-objective optimization scheduling module, a real-time regulation execution module, and a feedback correction module. The load data acquisition module is deployed at the main power circuits and key energy-consuming equipment within the building, continuously collecting signals of voltage, current, power factor, active power, reactive power, and equipment operating status, and performing data aggregation and normalization processing according to a preset time granularity. The short-term load forecasting module receives historical load sequences output by the load data acquisition module and integrates weather forecast data, building usage plan information, holiday identifiers, and time-series characteristic variables. It uses an improved time series analysis model to perform point-by-point load forecasting within the next 15 minutes to 4 hours, generating a high-resolution short-term load curve. This forecasting process introduces a sliding window adaptive weighting mechanism to dynamically adjust the contribution weight of historical data, improving the ability to capture sudden load changes.

[0007] Furthermore, the dynamic energy efficiency assessment module calculates the instantaneous energy efficiency index of the building as a whole and its zoned units based on short-term load forecast results and current system operating parameters. This module establishes an equipment-level energy efficiency benchmark library, storing standard energy consumption curves for various electrical devices under different load rates, and identifies inefficient operating sections by comparing the deviation between actual operating energy consumption and benchmark values ​​online. Simultaneously, this module combines ambient temperature and humidity, and occupant density sensing data to calculate the actual service efficiency of the air conditioning and ventilation systems, quantifying the electricity cost consumed per unit of service quality. The energy efficiency score output by the dynamic energy efficiency assessment module serves as one of the core inputs to the multi-objective optimization scheduling module.

[0008] Furthermore, the multi-objective optimization scheduling module comprehensively considers energy-saving objectives, power quality objectives, and equipment lifespan maintenance objectives, constructing a three-dimensional optimization function. The energy-saving objective focuses on minimizing the total electricity consumption within the predicted time period; the power quality objective is constrained by maintaining the bus voltage stable within ±5% of the rated range and controlling the harmonic distortion rate below 8%; the equipment lifespan maintenance objective is reflected by limiting the start-up and shutdown frequency and load fluctuation amplitude of key equipment such as chillers and transformers. This module adopts a hierarchical solution strategy. First, it generates regional scheduling sub-tasks based on the building's functional area division. Then, it uses a distributed iterative algorithm to solve the resource competition and coordination relationships between each sub-task, ultimately outputting a comprehensive scheduling scheme that includes equipment start-up and shutdown commands, power adjustment setpoints, and energy routing paths.

[0009] Furthermore, the real-time control execution module receives the scheduling scheme output by the multi-objective optimization scheduling module and parses it into control commands for specific actuators. This module supports multiple communication protocols and establishes bidirectional connections with frequency converters, intelligent circuit breakers, and building automation system controllers, ensuring reliable issuance of control commands and real-time feedback of execution status. For high-power equipment with inertial characteristics, the control execution module implements phased, gradual adjustments to avoid grid impact caused by abrupt power changes. Simultaneously, the module incorporates safety interlocking logic; when electrical parameters exceed limits, it automatically activates protection procedures to prioritize power supply safety.

[0010] Furthermore, the feedback correction module continuously monitors the actual load response after the control operation is executed, and compares and analyzes the measured data with the predicted values ​​and scheduling expectations. When the deviation exceeds the preset tolerance band, the feedback correction module triggers an online update mechanism for model parameters, fine-tuning the weight coefficients in the short-term load forecasting module and the baseline curve in the dynamic energy efficiency assessment module. This module also records the effect data of each control operation, forming a closed-loop learning sample set, which is used to periodically retrain the core algorithm model, thereby achieving continuous evolution of system performance.

[0011] Preferably, the short-term load forecasting module incorporates a load classification and identification mechanism. By performing wavelet packet decomposition on the original current waveform, energy distribution characteristics of each frequency band are extracted. Combined with equipment start-up and shutdown event markers, automatic identification of nonlinear loads, impulsive loads, and continuously operating loads is achieved. Dedicated forecasting sub-models are established for different types of loads, and differentiated confidence weights are assigned according to their fluctuation characteristics during the final forecast result fusion stage, thereby improving the overall forecast accuracy.

[0012] Preferably, the multi-objective optimization scheduling module integrates photovoltaic power generation forecast information and energy storage battery state-of-charge data. Under the premise of meeting local load demand, it prioritizes the scheduling of clean energy power supply and decides the timing of energy storage charging and discharging based on time-of-use pricing signals. When the external power grid enters peak pricing periods and the energy storage capacity is sufficient, the system automatically switches part of the load to off-grid operation mode to reduce electricity purchase costs.

[0013] Preferably, the dynamic energy efficiency assessment module sets up dynamic mapping rules for energy efficiency levels, converting energy efficiency scores over a continuous time period into visual energy efficiency labels, which are then pushed to the building operation and maintenance management platform. Maintenance personnel can locate high-energy-consuming areas based on the distribution of energy efficiency labels, assisting in developing plans for equipment upgrades or operational strategy adjustments.

[0014] Preferably, the feedback correction module is equipped with an abnormal operating condition identification unit, which uses statistical process control methods to monitor the morphological variation of the load curve. When a phenomenon of continuous deviation from the normal mode is detected, potential equipment fault warning information is generated, and preventive maintenance inspections are recommended.

[0015] Preferably, the system operates within a three-tiered time-scale collaborative framework, comprising an ultra-short-term control layer, a short-term scheduling layer, and a medium-to-long-term planning layer. The ultra-short-term control layer has a response cycle in the second range and is responsible for handling instantaneous load disturbances and voltage fluctuations; the short-term scheduling layer executes multi-objective optimized scheduling schemes with a basic scheduling interval of 15 minutes; and the medium-to-long-term planning layer generates equipment maintenance plans and energy-saving potential analysis reports based on historical operating data on a daily or weekly basis.

[0016] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0017] This system, by constructing a complete technology chain from load sensing and trend prediction to dynamic optimization and closed-loop control, achieves a fundamental shift in building power systems from passive response to proactive control. The short-term load forecasting module integrates multi-source heterogeneous information and employs an adaptive weighting mechanism, significantly improving the accuracy of predicting complex electricity consumption behaviors and providing a reliable basis for proactive energy-saving decisions. The dynamic energy efficiency assessment module breaks through the traditional single energy consumption metering model, introducing service efficiency conversion and equipment benchmark comparison, making energy efficiency evaluation more scientific and reasonable. The multi-objective optimization scheduling module incorporates energy saving, power quality, and equipment maintenance into a unified optimization framework, resolving the problem of conflicting control objectives and achieving synergistic improvement in multi-dimensional performance. The feedback correction module establishes a continuous model evolution mechanism, enabling the system to self-optimize as the operating environment changes. Overall, this invention effectively reduces building energy consumption by more than 12%, reduces the number of start-ups and shutdowns of critical equipment by 30%, extends the service life of electrical equipment, and improves power supply reliability and intelligent operation and maintenance, providing a system-level solution for modern green building energy management. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the overall technical architecture of the building power energy-saving control system based on load forecasting and dynamic regulation proposed in this invention. Detailed Implementation

[0019] Please refer to Figure 1 This invention relates to a building energy-saving control system based on load forecasting and dynamic regulation. Its technical architecture comprises a load data acquisition module, a short-term load forecasting module, a dynamic energy efficiency assessment module, a multi-objective optimization scheduling module, a real-time control execution module, and a feedback correction module. The system is deployed on a hybrid computing platform equipped with a central energy management server and distributed edge controllers. Its overall operation relies on a physical information fusion system composed of a high-precision sensor network, an industrial-grade communication bus, and localized data processing nodes. The core objective of this system is to achieve accurate perception, forward-looking prediction, multi-dimensional objective collaborative optimization, and closed-loop adaptive regulation of complex electrical loads within buildings, thereby maximizing energy utilization efficiency and extending the service life of critical equipment while ensuring power supply security and service quality.

[0020] The system's operation begins with the load data acquisition module. As the sensing front end of the entire control system, this module undertakes the fundamental function of constructing high-quality raw datasets. By deploying three-phase smart meters, Hall effect current sensors, multi-functional power analyzers, and equipment status monitoring terminals in various circuits of the building's power distribution system, the module synchronously acquires the full range of electrical and non-electrical parameters, including voltage, current, active power, reactive power, power factor, frequency, harmonic content, circuit breaker position, and contactor operation signals. The sampling frequency is differentiated according to the load type: for high-power periodic loads such as air conditioning units and elevator traction machines, the sampling interval is 200 milliseconds; for low-dynamic-characteristic branches such as lighting circuits and socket loads, the sampling interval is set to 5 seconds; and for the main incoming line cabinet of the bus section, a high-speed sampling mechanism of once every 100 milliseconds is used to capture transient disturbances. All acquisition units are equipped with time synchronization chips, achieving microsecond-level clock alignment through the IEEE 1588 precision time protocol to ensure time consistency of data across regions.

[0021] The collected raw data is first preprocessed at the local edge gateway. The preprocessing includes bad pixel removal, zero drift correction, spike filtering, and normalization transformation. Bad pixel identification uses a sliding window median comparison method; when a data point deviates from the median of its five nearest neighbors by more than three times the standard deviation, it is considered an outlier and replaced with linear interpolation. Zero drift correction addresses sensor baseline offset caused by long-term operation by automatically collecting idle data during the low-load period in the early morning each day and updating the baseline compensation parameters. Spike filtering uses an improved Savitzky-Golay smoothing algorithm, effectively suppressing high-frequency noise interference while preserving the main signal characteristics. Normalization transformation maps physical quantities of different dimensions to a unified numerical range, specifically using a minimum-maximum scaling formula.

[0022]

[0023] in, Represents the original measurement value. and These are the historical extreme values ​​of the variable over the past 7 days. This is the normalized output value. This operation ensures that subsequent models treat input variables with different amplitude ranges equally, avoiding weight bias caused by differences in numerical scale. The preprocessed data is aggregated into a statistical series with a basic time granularity of 15 minutes, including the mean, peak value, variance, and rate of change. After being tagged with timestamps, it is uploaded to the central database for use by subsequent modules.

[0024] The short-term load forecasting module receives historical load sequences from the load data acquisition module and integrates external auxiliary information to extrapolate future electricity demand trends. The core design principle of this module is to enhance its responsiveness to sudden event-driven load fluctuations, particularly maintaining forecast stability during scenarios such as concentrated meetings, large-scale events, or extreme weather events. Input data includes: total active power sequences over the past 72 hours, 15-minute intervals; independent load curves for each functional zone; outdoor dry-bulb temperature and relative humidity forecasts; building occupancy plans (such as meeting room reservations and performance schedules); and holiday identifiers and time series codes (hours, days of the week, and months). All input variables have undergone the aforementioned normalization process.

[0025] The prediction model employs a deep neural network structure based on gated recurrent units (GRUs), specifically a two-layer stacked GRU network, with each layer containing 64 hidden units. An attention mechanism layer is integrated at the top of the network to dynamically allocate weight contributions for different historical time steps. (Attention weights) The calculation method is as follows:

[0026]

[0027] in, Indicates the encoder at time... The hidden state, The current context vector of the decoder. For additive scoring functions, defined as follows: , , , These are trainable parameters. This mechanism allows the model to automatically focus on electricity consumption patterns during the same time period of the previous day when predicting peak loads, or to increase the weighting of meteorological factors during cold waves.

[0028] The model training employs a sliding window adaptive weighting strategy. Traditional fixed-weight methods are prone to hysteresis errors when facing structural changes; therefore, this system introduces an exponential decay factor. This assigns higher training weights to historical data that is more recent than the current time. Let the window length be... Then the first The weights of each historical sample are ,in The value is set to 0.98. This parameter was determined through grid search combined with cross-validation, achieving the best root mean square error performance on the test set. During offline training, the model used a complete dataset from the past year, containing 35,040 15-minute time slices. During online inference, the model updates its predictions every 15 minutes, outputting point-by-point load predictions for the next 16 time steps (i.e., the next 4 hours), maintaining a resolution at the 15-minute level.

[0029] To further improve prediction accuracy, the module incorporates a load classification and identification subsystem. This subsystem extracts multi-band energy features by performing wavelet packet decomposition on the main circuit current waveform. Using the Daubechies wavelet basis db4, the current signal with a sampling frequency of 50 Hz is decomposed into 16 sub-bands through a 4-level process. The energy proportion of each sub-band is then calculated. ,in For the first A sequence of wavelet coefficients for each sub-band was used. Simultaneously, a supervised label set was established based on equipment start-up and shutdown event logs to annotate waveform fingerprints corresponding to typical load types. A random forest classifier was used to train a recognition model to distinguish between nonlinear loads (such as variable frequency air conditioners and LED driver power supplies), impulsive loads (such as elevator starts and welding machines), and continuously operating loads (such as data center servers and chilled water pumps). Dedicated GRU prediction sub-models were constructed for each of the three load types, and differentiated confidence weights were assigned during the final fusion stage: nonlinear loads received a weight of 0.6 due to their high volatility requiring greater attention; impulsive loads received a weight of 0.8, considering their sudden but identifiable characteristics; and continuously operating loads received a weight of 0.4 due to their stable trend. The final prediction result was a weighted average output, significantly reducing the risk of misjudgment by a single model.

[0030] The dynamic energy efficiency assessment module quantifies the instantaneous energy efficiency levels of the entire building and its sub-areas based on short-term load forecasts and current system operation. This module breaks through the limitations of traditional energy consumption measurement based solely on kilowatt-hours by introducing the concept of "service equivalent energy consumption," linking electricity consumption to the actual use value provided. Internally, the module maintains an equipment-level energy efficiency benchmark library, storing the rated energy efficiency curves of various major electrical equipment under standard operating conditions. The benchmark data for chiller units includes the coefficient of performance (COP) under different combinations of chilled water outlet temperature, cooling water inlet temperature, and load rate; centrifugal fans record the three-dimensional relationship surface of airflow-head-power; and LED lighting fixtures record the correspondence between luminous flux and input power. All benchmark data originates from equipment factory test reports, third-party certification documents, and on-site commissioning test results, and is formatted and stored in an embedded SQLite database.

[0031] The evaluation process is divided into two levels: the equipment level and the system level. At the equipment level, the module reads current operating parameters in real time, calculates the actual energy efficiency index, and compares it with the benchmark value. For example, for a running screw chiller unit, the evaporator inlet and outlet water temperature difference, flow rate, compressor current, and input power are collected to calculate the ratio of actual cooling capacity to power consumption, obtaining the real-time COP. If this value is lower than 90% of the benchmark COP under the same operating conditions, it is marked as an inefficient operating segment, and an energy efficiency deviation alarm is triggered. The degree of deviation is expressed as a percentage and is defined as follows: ,when Initiate the in-depth diagnostic process when the threshold is >15%.

[0032] At the system level, the module integrates environmental perception data to calculate the overall efficiency of the air conditioning and ventilation systems. Wireless temperature and humidity sensors deployed in various office areas, lobbies, and corridors report environmental parameters every 2 minutes. Personnel density is estimated jointly using Wi-Fi probes and a video analytics system. The module establishes a thermal comfort evaluation model, using the PMV-PPD index system to calculate current indoor thermal environment satisfaction. The electrical energy consumed per unit of satisfied person per hour is defined as "service equivalent energy consumption," expressed as: ,in This represents the total power consumption of the air conditioning system during the specified time period. To represent the average number of people present, For the length of time, This is the thermal comfort adjustment efficiency factor (ranging from 0.7 to 1.2, reflecting the deviation between the actual perceived temperature and the set temperature). This indicator allows maintenance personnel to compare the actual service costs of air conditioning systems across different seasons and under different traffic conditions.

[0033] The evaluation results are ultimately output as an energy efficiency score, ranging from 0 to 100. The scoring rules employ a multi-factor weighted composite method: a base score of 100 points, with deductions based on equipment energy efficiency deviations, deducting 0.5 points for every 1% average energy efficiency loss; adjustments are made based on the ratio of service equivalent energy consumption to the historical average for the same period, with additional deductions for deviations exceeding 110%; if new equipment not included in the benchmark database is put into operation, an initial bonus is awarded based on its design energy efficiency level, up to a maximum of 5 points. The score is refreshed every 15 minutes and used as a key input to the multi-objective optimization scheduling module.

[0034] The multi-objective optimization scheduling module is the decision-making center of this system, responsible for generating globally optimal energy dispatch instructions under multiple constraints. The module receives energy efficiency scores from the dynamic energy efficiency assessment module, future load curves from the short-term load forecasting module, photovoltaic power generation forecast data, energy storage battery state of charge, and time-of-use pricing signals, comprehensively forming a multi-dimensional optimization problem. The optimization objectives cover three incommensurable dimensions: energy efficiency, power quality, and equipment lifespan maintenance.

[0035] The energy-saving target focuses on minimizing the total electricity consumption of the building in the next 4 hours, which can be mathematically expressed as: ,in For the first Net power drawn from the grid in a 15-minute time period This refers to a time interval. Power quality targets are embodied in two sets of hard constraints: the bus voltage must be maintained within ±5% of its rated value, i.e. The total harmonic distortion (THD) must be less than 8%, i.e., THD(t) < 8%. Equipment life maintenance targets are achieved by restricting operational behavior: the number of start-ups and shutdowns of chiller units shall not exceed 4 times per day, and the interval between adjacent start-ups and shutdowns shall not be less than 90 minutes; the change rate of transformer load rate is limited to no more than 5 percentage points per minute to prevent fatigue aging of insulation materials.

[0036] To solve this multi-objective nonlinear programming problem, the module employs a hierarchical optimization strategy. The first layer is spatial decomposition, dividing the overall load into several scheduling regions based on the building's functional layout, such as office areas, commercial podiums, underground parking, and data centers. Each region is treated as an independent resource agent with local optimization capabilities. The second layer is time coordination, dividing the next 4 hours into 16 discrete time periods and constructing a dynamic programming framework. The third layer is algorithm execution, using an improved non-dominated sorting genetic algorithm (NSGA-II) for Pareto front search. The population size is set to 100, with a crossover probability of 0.8, a mutation probability of 0.1, and 50 generations. The algorithm outputs a set of non-dominated solutions, representing feasible scheduling schemes under different emphasis directions.

[0037] In the final decision-making stage, a preference selection mechanism is introduced. The system presets priority weights: during periods of flat electricity pricing, energy saving accounts for 60%, power quality for 25%, and equipment lifespan for 15%; during periods of peak electricity pricing, energy saving weight increases to 80%, while the other two decrease to 10% each; during periods of major event security, power quality weight is increased to 70% to ensure absolute reliability. Based on the current time period type, a corresponding weight vector is loaded, and the Pareto solution set is weighted and scored. The scheme with the highest comprehensive score is selected as the final dispatch instruction. Outputs include: target set temperature for air conditioning systems in each area, fresh air valve opening instructions, target speed of water pump inverters, dimming percentage of lighting circuits, charging and discharging power instructions for energy storage systems, and diesel generator standby status indicators.

[0038] The real-time control and execution module is responsible for translating high-level scheduling instructions into operational commands recognizable by the lower-level actuators. This module runs on an industrial-grade programmable logic controller (PLC) platform, featuring millisecond-level response capabilities and a highly reliable redundant design. Internally configured with a protocol conversion engine, it supports multiple communication standards such as Modbus TCP, BACnet IP, KNX, and MQTT, enabling seamless integration with frequency converters, intelligent circuit breakers, building automation DDC controllers, and energy storage inverters. All control links employ encrypted tunnel transmission, utilizing the AES-256 algorithm to ensure command integrity and tamper-proof protection.

[0039] For high-power equipment with significant inertia, the execution module implements a gradual adjustment strategy. Taking a central air conditioning chiller unit as an example, when a cooling command is received, the supply water temperature setpoint is not directly jumped to the target value. Instead, it is broken down into multiple intermediate steps, with each step adjusting no more than 1 degree Celsius and the interval between adjacent steps being no less than 5 minutes. This process is achieved through a built-in state machine, and the state transition condition is that the previous level of adjustment has been stable for more than 3 minutes and the system pressure fluctuation is less than 5%. Similarly, the lighting system dimming adopts a ramp transition, with the brightness change rate limited to no more than 5% per second to avoid causing visual discomfort.

[0040] The module incorporates multiple safety interlocking logics. When any feeder current exceeds 110% of the circuit breaker's rated value for 10 seconds, a tiered unloading procedure is automatically initiated, prioritizing the disconnection of non-critical loads such as advertising light boxes and decorative lighting. If the voltage drops below 90%, all new load connections are immediately blocked, and a startup request is sent to the emergency power system. All protection actions generate event logs, including the occurrence time, triggering conditions, executed actions, and recovery status, stored in an independent safety audit database for at least 3 years.

[0041] The feedback correction module forms the final link in the system's closed-loop control, responsible for model parameter correction and performance iterative optimization. This module continuously collects actual load response data after control measures are implemented and performs residual analysis against short-term load forecasts and multi-objective optimal scheduling expectations. A comprehensive deviation index is defined. ,in , , These represent the actual-to-expected deviations of active power, voltage amplitude, and frequency, respectively. , , These are weighting coefficients, set to 1.0, 0.5, and 0.3 respectively, to reflect the differences in importance among the various indicators. When When the sampling period exceeds the preset tolerance band (set to 7% of full scale) for three consecutive sampling periods, the online update mechanism of model parameters is triggered.

[0042] Parameter updates cover both the short-term load forecasting module and the dynamic energy efficiency assessment module. For the forecasting module, the weight matrix of the last fully connected layer of the GRU network is updated using the recursive least squares (RLS) method. Let the error signal... Then the weight update law is ,in The gain vector is calculated recursively from the covariance matrix. This method can achieve rapid adaptation without retraining the entire deep network, with a computation time of less than 200 milliseconds. For the energy efficiency assessment module, when a device is found to be in an inefficient segment identified by the system for an extended period, a baseline curve self-learning program is initiated. Measured data from all operating conditions of the device over the past 30 days are collected, and its true energy efficiency distribution is reconstructed using kernel density estimation. Based on this, the original manufacturer's baseline value is fine-tuned, with the deviation correction not exceeding ±10% to prevent overfitting.

[0043] In addition, the feedback correction module is equipped with an abnormal operating condition identification unit, using statistical process control (SPC) technology to monitor the evolution of the load curve shape. It calculates the fourth-order moment characteristics of the hourly load sequence: mean, standard deviation, skewness, and kurtosis, and projects them into a four-dimensional feature space. Under normal operating conditions, these characteristics fluctuate slightly around the steady-state center point. The module establishes a Mahalanobis distance threshold model; when the Mahalanobis distance from the real-time observation point to the historical center exceeds the chi-square distribution critical value (4 degrees of freedom, significance level 0.01), it is determined to be an abnormal mode. Possible causes include increased vibration due to motor bearing wear, increased contact resistance due to loose cable joints, or command drift caused by control system errors. At this time, potential equipment fault warning information is generated, and preventative maintenance operations such as infrared thermography, partial discharge detection, or control loop inspection are recommended.

[0044] The system operates within a three-tiered time-scale collaborative framework. The ultra-short-term control layer, with a response cycle of 1 to 10 seconds, focuses on suppressing instantaneous voltage flicker, compensating for reactive power fluctuations, and clearing short-circuit faults. This is accomplished collaboratively by the PLC's built-in PID controller and the Static Var Generator (SVG). The short-term scheduling layer operates on a 15-minute basic scheduling cycle, executing the aforementioned multi-objective optimization scheduling scheme, coordinated by the central server. The medium- and long-term planning layer operates on a daily or weekly cycle, generating two types of outputs based on accumulated historical operating data: first, equipment maintenance plans, which predict remaining service life and schedule preventative maintenance windows based on the cumulative number of chiller start-ups and shutdowns, transformer hotspot temperature integrals, and circuit breaker mechanical operation counts; second, energy-saving potential analysis reports, which identify structural energy-saving opportunities, such as building envelope modifications, replacement of high-efficiency equipment, or operational strategy optimization suggestions, by comparing energy consumption intensity per unit area under different seasons and usage patterns.

[0045] Preferably, the multi-objective optimization scheduling module integrates distributed photovoltaic (PV) power generation capacity prediction information. PV prediction is based on cloud cover, solar radiation intensity, and temperature data provided by numerical weather prediction (NWP), combined with array orientation, tilt angle, and attenuation coefficient. A fusion of physical modeling and machine learning methods is used to predict power output for the next four hours. The energy storage battery management system uploads state of charge (SOC) and state of health (SOH) data in real time. When formulating plans, the scheduling module prioritizes clean energy supply. If there is still a surplus after meeting local load demand, it controls the bidirectional converter to feed power back to the grid. When entering peak electricity price periods and the SOC exceeds 30%, it automatically switches some secondary loads to off-grid operation mode, where they are independently powered by the energy storage system, significantly reducing electricity purchase costs. Economic calculations show that under typical industrial and commercial electricity price structures, this strategy can save more than 18% of electricity costs annually.

[0046] Preferably, the dynamic energy efficiency assessment module sets dynamic mapping rules for energy efficiency levels. The average energy efficiency score over a continuous 24-hour period is divided into five levels: 90 points or above is "Excellent," 80-89 points is "Good," 70-79 points is "Satisfactory," 60-69 points is "Needs Improvement," and below 60 points is "Alarm." The level results are pushed to the building operation and maintenance management platform via a RESTful API, visually displaying the energy efficiency status of each area on an electronic map in the form of a color heatmap. Operation and maintenance personnel can click on any area to view detailed diagnostic information, including a list of major energy-consuming equipment, sources of energy efficiency deviations, and a list of improvement suggestions. This function greatly improves management transparency and decision-making efficiency.

[0047] This embodiment constructs a building energy-saving control system with self-sensing, predictive inference, intelligent decision-making, precise execution, and continuous evolution through the organic coupling of the above six modules. The system achieves a fundamental shift from passive response to active regulation, overcoming the technical bottlenecks of fragmented objectives, delayed response, and insufficient adaptability in traditional control strategies. Experimental data shows that in typical office building application scenarios, after the system was put into operation, overall energy consumption decreased by 13.7%, the annual start-up and shutdown frequency of chiller units decreased by 35%, the bus voltage qualification rate increased from 91.2% to 98.6%, and the harmonic distortion rate was stably controlled within 6.8%. The system not only creates significant economic benefits but also promotes the development of building energy management towards intelligence, refinement, and sustainability.

[0048] Existing building power control systems generally rely on fixed threshold comparisons and simple logic interlocks to start and stop equipment, lacking the ability to predict future load changes. Such systems often experience response delays when dealing with sudden increases in power equipment or surges in air conditioning load caused by weather changes, leading to voltage dips or transformer overloads. Even worse, conservative operating strategies adopted to mitigate risks result in equipment operating at low load rates for extended periods, leading to energy waste due to over-utilization. This invention introduces a high-precision short-term load forecasting module to identify potential load change trends in advance, providing a valuable time window for proactive regulation. The forecast results drive a multi-objective optimization scheduling module to generate a comprehensive strategy that considers energy saving, power quality, and equipment lifespan, fundamentally changing the fragmented management model of "treating symptoms rather than the root cause."

[0049] Compared to existing technologies, the key difference of this invention lies in establishing a complete closed-loop control chain of "perception-prediction-decision-execution-feedback". Traditional systems often remain in an open-loop or semi-closed-loop state, unable to correct model parameters based on actual control effects. The feedback correction module of this invention not only monitors deviations but also drives the online updating and periodic retraining of the core algorithm model, enabling the system to self-adjust as building usage patterns evolve and equipment aging progresses. This mechanism ensures the stability and superiority of the system's long-term performance, avoiding the problem of control effect attenuation due to model degradation.

[0050] Furthermore, this invention achieves significant innovation at the multi-objective optimization scheduling level. Existing technologies often treat energy saving as the sole objective, neglecting the potential power quality problems or accelerated equipment wear and tear that may result. For example, simply pursuing lowering the air conditioner's set temperature to reduce operating time may lead to frequent compressor start-stop cycles, increasing energy consumption per unit cooling capacity and shortening its lifespan. This invention incorporates three objectives into a unified mathematical framework, obtains a balanced solution set through Pareto optimality, and dynamically adjusts priority weights based on the operating scenario, achieving a synergistic improvement in multi-dimensional performance. This system-level thinking transcends the limitations of local optimization, truly achieving the overall optimal operation of the building's energy system.

[0051] The introduction of a dynamic energy efficiency assessment module also represents a significant advancement. Traditional energy management systems only provide meter readings and sub-metering reports, failing to answer the core question of "how much electricity was consumed and what service was received." This invention's pioneering service equivalent energy consumption index links energy consumption with service quality parameters such as personnel comfort and space utilization intensity, giving energy efficiency evaluation practical business significance. Maintenance personnel no longer rely solely on intuition to judge energy-saving effectiveness but instead use scientifically quantified indicators to guide renovation investments and operational optimization, greatly improving the accuracy and persuasiveness of management decisions.

[0052] In summary, the technical solution disclosed in this embodiment, through the deep integration of advanced sensing, artificial intelligence prediction, multi-objective optimization, and closed-loop feedback control technologies, constructs a power-saving control system suitable for modern complex building environments. The system not only achieves breakthroughs in individual technical indicators but also realizes a paradigm shift at the overall architectural level.

Claims

1. A building power saving control system based on load prediction and dynamic regulation, characterized in that, The application relates to a building energy management system, comprising: a load data acquisition module for continuously collecting electrical parameters and equipment operation state signals from main power circuits and key energy-consuming equipment inside a building to obtain an original load data set; a short-term load prediction module for receiving the original load data set, fusing meteorological prediction data, building use plan information, holiday identification and time sequence characteristic variables, and performing point-by-point load prediction in a future period by using an improved time sequence analysis model to generate a short-term load curve, wherein the short-term load prediction module comprises a sliding window self-adaptive weighting unit for dynamically adjusting the contribution weight of historical load data in the prediction process to improve the capturing ability for sudden load changes; a dynamic energy efficiency evaluation module for calculating the instantaneous energy efficiency indexes of the whole building and partition units based on the short-term load curve and current system operation parameters, wherein the dynamic energy efficiency evaluation module comprises an equipment-level energy efficiency benchmark library for storing standard energy consumption curves of various power-consuming equipment under different load rates; and a service performance conversion unit for converting the actual service performance of air conditioning and ventilation systems in combination with environmental temperature and humidity and personnel density sensing data to quantify the power consumption cost of unit service quality; a multi-objective optimization scheduling module for comprehensively considering energy-saving targets, power supply quality targets and equipment life maintenance targets, constructing a three-dimensional optimization function, and generating a comprehensive scheduling scheme based on the instantaneous energy efficiency indexes and the short-term load curve, wherein the multi-objective optimization scheduling module comprises a hierarchical solving unit for generating regional scheduling subtasks according to building functional area division and solving the resource competition and cooperation relationship among the subtasks by using a distributed iterative algorithm; a real-time regulation and control execution module for receiving the comprehensive scheduling scheme and analyzing the same into control commands for specific execution mechanisms, wherein the real-time regulation and control execution module comprises a gradual adjustment unit for implementing stage-by-stage gradual adjustment on high-power equipment with inertia characteristics to avoid power step change and cause power grid impact; a feedback correction module for continuously monitoring the actual load response after regulation and control execution, comparing measured data with predicted values and scheduling expected values, and triggering an online model parameter updating mechanism to finely adjust the weight coefficients in the short-term load prediction module and the benchmark curves in the dynamic energy efficiency evaluation module when the deviation exceeds a preset tolerance band.

2. The building power saving control system based on load prediction and dynamic regulation according to claim 1, characterized in that, The electrical parameters include voltage, current, power factor, active power and reactive power. 3.The building power saving control system based on load prediction and dynamic regulation of claim 1, wherein, The short-term load prediction module further comprises: a load classification and identification unit for automatically identifying nonlinear loads, impact loads and continuously running loads by wavelet packet decomposition of original current waveforms to extract frequency band energy distribution characteristics and in combination with equipment start-stop event markers; a special prediction submodel unit for respectively establishing special prediction submodels for the nonlinear loads, impact loads and continuously running loads; a differentiated confidence weight fusion unit for giving differentiated confidence weights according to the fluctuation characteristics of various loads in the final prediction result fusion stage to improve overall prediction accuracy.

4. The building power saving control system based on load prediction and dynamic regulation according to claim 1, characterized in that, The multi-objective optimization scheduling module is further configured to integrate photovoltaic power generation prediction information and energy storage battery state of charge data, prioritize scheduling clean energy power supply under the premise of meeting local load demand, and decide energy storage charging and discharging time according to a time-of-use electricity price signal, automatically switch part of the load to an off-grid operation mode when the external power grid enters a peak electricity price period and the energy storage capacity is sufficient.

5. The building power saving control system based on load forecasting and dynamic regulation according to claim 1, characterized in that, The dynamic energy efficiency evaluation module further comprises: An energy efficiency grade dynamic mapping unit configured to convert energy efficiency scores in consecutive time periods into visual energy efficiency labels; An energy efficiency label pushing unit configured to push the visual energy efficiency labels to a building operation and maintenance management platform for operation and maintenance personnel to locate high energy consumption areas. 6.The building power saving control system based on load prediction and dynamic regulation according to claim 1, characterized in that, The feedback correction module further comprises: An abnormal working condition identification unit configured to monitor the shape variation of a load curve using a statistical process control method, and generate device potential fault warning information when detecting a phenomenon of continuously deviating from a normal mode.

7. The building power saving control system based on load forecasting and dynamic regulation according to claim 3, characterized in that, The load classification identification unit comprises: A wavelet packet decomposition subunit configured to implement wavelet packet decomposition on the original current waveform to generate a plurality of sub-bands; An energy proportion calculation subunit configured to calculate the energy proportion of each sub-band to form an energy distribution feature vector; A supervised classification subunit configured to combine device start-stop event logs as supervised labels, train a classification model to distinguish the nonlinear load, the impact load and the continuous operation load. 8.The building power saving control system based on load prediction and dynamic regulation according to claim 1, characterized in that, The real-time regulation execution module is further configured to support multiple communication protocols, and establish a bidirectional connection with a variable frequency drive, an intelligent circuit breaker and a building automation system controller to ensure reliable issuance of control instructions and real-time return of execution status. 9.The building power saving control system based on load forecasting and dynamic regulation of claim 1, wherein, The system operates in a three-level time scale collaborative framework, including: an ultra-short-term regulation layer configured to respond to transient load disturbance and voltage fluctuation with a response period of seconds; a short-term scheduling layer configured to execute a comprehensive scheduling scheme generated by the multi-objective optimization scheduling module with fifteen minutes as a basic scheduling interval; and a medium and long-term planning layer configured to generate a device maintenance plan and an energy saving potential analysis report based on historical operation data on a daily or weekly cycle.

10. The building power saving control system based on load forecasting and dynamic regulation according to claim 1, characterized in that, The three-dimensional optimization function in the multi-objective optimization scheduling module has an energy saving target of minimizing total power consumption in a prediction period as the core, a power supply quality target of maintaining bus voltage stability and controlling harmonic distortion rate as a constraint condition, and a device life maintenance target embodied by limiting the start-stop frequency and load mutation amplitude of key devices.

Images