AI-based thermal load prediction and multi-heat-source optimization scheduling system

Abstract

The invention relates to the technical field of general control or regulation systems, and discloses an AI-based thermal load prediction and multi-heat-source optimization scheduling system, which comprises a source characteristic storage unit, which is used for dividing heat source equipment into an inertial base load group and an agile regulation group according to a response time constant; the load trend prediction unit is used for calculating a thermal load change slope by adopting a time sequence prediction model; the control instruction generation unit is used for decoupling the total heat load demand by utilizing a low-pass filtering algorithm according to the change slope; the collaborative scheduling unit is used for establishing a frequency division mapping mechanism, mapping a low-frequency base quantity demand to an inertial base load group and mapping a high-frequency increment demand to an agile adjustment group; and the feedback correction unit is used for generating a correction compensation amount and only superposing the correction compensation amount to a control instruction for the agile adjustment group when the change rate of the total return water temperature of the heat supply network exceeds the limit, and adjustment oscillation caused by heat supply network lag is avoided through frequency division decoupling and a targeted feedback framework.

Description

A heat load prediction and multi-heat source optimization scheduling system based on AI Technical Field

[0001] This invention relates to an AI-based heat load prediction and multi-heat source optimization scheduling system, belonging to the general field of control or regulation system technology. Background Technology

[0002] In complex industrial control systems that couple centralized heating and distributed energy, the controlled objects exhibit large inertia, long time lag, and nonlinear physical characteristics. Traditional control strategies, when dealing with load signals under varying operating conditions, lack the dynamic perception capability of the time-frequency characteristics of disturbance signals. This easily leads to phase mismatch between the control system's adjustment commands and the response characteristics of physical actuators, causing system oscillations or static regulation saturation. To address the matching problem between supply-side heterogeneous actuators with different response characteristics and costs and demand-side dynamic loads, existing technologies generally adopt a feedforward-feedback composite control strategy based on frequency domain decomposition. The general logic pre-sets a fixed filter cutoff frequency, decoupling the total load demand into low-frequency base load trends and high-frequency transient disturbances, which are then mapped to the inertial base load loop and the agile adjustment loop, respectively. This attempts to achieve a slow source to slow load and a fast source to fast load. Static division of labor for fast loads achieves stable system operation and cost control. However, in actual variable operating conditions, the control architecture based on static frequency domain division faces severe physical and logical challenges. The load disturbance signal of the heating network system does not have constant spectral properties in the time dimension. The initial stage of sudden step disturbance is manifested as high-frequency increment. If the disturbance persists over time, its physical properties gradually transform into steady-state DC components. Existing control systems lack the ability to dynamically perceive and adapt to the evolution of the time-frequency characteristics of disturbance signals, resulting in a serious phase mismatch between control logic and physical facts. This makes it impossible for the system to dynamically reconstruct the control topology for the non-stationary time-frequency characteristics of the load signal. When the high-frequency disturbance transforms into a steady-state load, the control system still locks it in the agile adjustment loop, forcing high-cost fast-response equipment to bear the DC load that should be borne by the base load equipment for a long time.

[0003] Other solutions in the industry that attempt to solve the dynamic scheduling problem also have limitations in their approach. For example, Chinese invention patent CN104791903B discloses a heating network intelligent scheduling system that adopts a load forecasting-network balancing technical path. It determines the global dynamic balance control scheme by combining a load forecasting unit with real-time hydraulic and thermal analysis. The limitation of this method is that it still treats the total load as a whole for global matching. The control architecture does not decouple the load characteristics in the frequency domain and lacks a mechanism to distinguish between transient high-frequency disturbances and steady-state low-frequency base loads. It also cannot solve the aforementioned load spectrum drift problem and lacks a self-learning control mechanism with the ability to dynamically transfer energy references. It is difficult to automatically migrate loads that have evolved into a steady state from high-cost agile heat sources to low-cost inertial heat sources.

[0004] Therefore, the technical problem to be solved by this invention is how to construct a system that can sense the time-varying evolution of load spectrum characteristics in real time, ensure the stable operation of the system, and automatically drive control commands to dynamically migrate and rebalance between actuator loops with different response characteristics in accordance with energy efficiency logic. Summary of the Invention

[0005] To address the problems mentioned in the background art, the technical solution of this invention is as follows: A heat load prediction and multi-heat source optimization scheduling system based on AI, comprising: a source characteristic storage unit for storing the response time constants of each heat source device in the heating network, and dividing the heat source devices into an inertial base load group and an agile adjustment group based on the response time constants, wherein the inertial base load group has a response time constant greater than a preset benchmark value, and the agile adjustment group has a response time constant less than a preset benchmark value; a load trend prediction unit for calculating the slope of heat load change in the heating area within the future pipeline transmission lag period based on historical operating data and environmental parameters of the heating area using a time series prediction model; and a control command generation unit for solving the real-time total heat load demand using a low-pass filtering algorithm based on the slope of heat load change. The system is coupled with low-frequency base load demand and high-frequency incremental demand. The collaborative scheduling unit is used to establish a frequency-division mapping mechanism, which maps low-frequency base load demand to a first control command for the inertial base load group and high-frequency incremental demand to a second control command for the agile regulation group. The rate of change of the first control command is limited to the allowable ramp rate of the inertial base load group. The feedback correction unit is used to collect the rate of change of the total return water temperature of the heating network in real time. The collaborative scheduling unit is also used to execute targeted compensation logic. When the rate of change of the total return water temperature of the heating network exceeds the preset stable dead zone range, a correction compensation amount is generated and superimposed only on the second control command to adjust the output power of the agile regulation group while keeping the first control command unchanged, thereby using the agile regulation group to eliminate high-frequency thermal hysteresis fluctuations.

[0006] Preferably, the load trend prediction unit is also used to perform feedforward time window adaptive adjustment, the adjustment including: the flow-time mapping unit is used to collect the total circulating flow data of the heating network in real time, and call the preset flow and lag time nonlinear mapping relationship table to find the equivalent transmission lag time corresponding to the total circulating flow data; the load trend prediction unit sets the equivalent transmission lag time as the effective prediction step size of the time series prediction model, so that the prediction time span of the heat load change slope is kept synchronized with the current hydraulic transport speed of the network in real time.

[0007] Preferably, the collaborative scheduling unit is also used to execute energy reference dynamic migration logic, the logic including: the reference migration arbitration unit is used to calculate the integral average of the second control command within a preset sliding time window; when the integral average continuously exceeds a preset dwell threshold, the collaborative scheduling unit generates a reference migration compensation amount, the rate of change of the reference migration compensation amount is limited by the maximum allowable ramp rate of the inertial base load group; the collaborative scheduling unit adds the reference migration compensation amount to the first control command, and simultaneously subtracts a control amount equal to the value of the reference migration compensation amount from the second control command, until the integral average returns to the preset balance dead zone range.

[0008] Preferably, the sliding time window executed by the reference migration arbitration unit is longer than the pipeline transmission lag period, and the dwell threshold is set as the energy determination boundary to distinguish between transient disturbances and steady-state load drift.

[0009] Preferably, the system further includes an adaptive performance correction unit for compensating for the performance degradation of the physical actuator. The adaptive performance correction unit is configured to: calculate the theoretical expected response of the agile adjustment group based on the second control command and the preset initial equipment gain coefficient; perform differential calculation between the theoretical expected response and the actual measured value of the rate of change of the total return water temperature of the heating network to obtain the steady-state residual value; when the steady-state residual value exceeds the preset fault judgment threshold, update the initial equipment gain coefficient in reverse according to the steady-state residual value, and use the updated gain coefficient to perform weighted correction on the second control command.

[0010] Preferably, the performance adaptive correction unit updates the initial device gain coefficient using the following logic: ,in, This is the updated gain coefficient. The gain coefficient currently in use. The preset correction step size factor, For steady-state residual values, This is the reference value for the rated response of the agile adjustment group.

[0011] Preferably, the inertial base load group includes ground source heat pump units, industrial waste heat recovery devices, or biomass boilers; the agile adjustment group includes gas-fired boilers or electrode boilers.

[0012] Preferably, the long short-term memory network model or gated cyclic unit network model used in the load trend prediction unit is deployed in the edge computing node adjacent to the heating area. The edge computing node directly obtains historical operating data and environmental parameters through the industrial fieldbus.

[0013] Preferably, the cutoff frequency of the low-pass filtering algorithm used by the control command generation unit is dynamically adjustable. The control command generation unit adjusts the cutoff frequency in real time according to the remaining adjustable capacity of the agile adjustment group. When the remaining adjustable capacity is lower than the preset safety value, the cutoff frequency is reduced to increase the proportion of low-frequency base quantity demand allocated to the first control command.

[0014] Preferably, the rate of change of the total return water temperature of the heating network collected by the feedback correction unit is obtained by performing differential calculation on the return water temperature measurement values ​​of multiple consecutive sampling cycles and then smoothing and filtering them. The preset stable dead zone range is determined based on the minimum adjustable power step size of the agile adjustment group.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. Based on the integral migration dynamic load spectrum adaptation and resource dynamic optimization collaborative scheduling unit, the present invention introduces a mechanism for monitoring the integral mean of agile adjustment commands and a benchmark dynamic migration mechanism to solve the problem of resource mismatch of actuators caused by the time drift of the load disturbance signal spectrum characteristics in multi-heat source control systems. When the high-frequency disturbance component exists continuously in the time dimension and is transformed into a steady-state DC component, the integral feedback logic generates a migration compensation amount limited by the ramp rate of the base load group. Without causing pipeline oscillation, the steady-state load that has been stuck on the high-cost agile heat source for a long time is smoothly transferred to the low-cost inertial heat source. The mechanism maintains the system's rapid response capability synchronously, realizes the dynamic reconstruction of the control topology as the load characteristics change, and ensures that heat sources with different response characteristics always work in the frequency domain range that matches the physical properties. This avoids the agile actuator being in a non-optimal integral saturation or high energy consumption state for a long time due to the static frequency division strategy.

[0016] 2. Based on flow-time mapping and variable time-delay feedforward phase synchronization, the flow rate and time coupling characteristics in fluid transport are utilized to construct a dynamic mapping relationship between the feedforward prediction time window and the real-time pipeline circulation flow. Addressing the nonlinear change in heat transfer lag time under variable flow conditions, the effective step size and filtering parameters of the time series prediction model are adjusted in real time to ensure strict phase synchronization between the feedforward control signal issuance time and the physical time when the heat wavefront arrives at the end. Based on the physical transmission mechanism, time-domain calibration eliminates the lead or lag errors that inevitably occur in fixed-time-window predictions in variable flow velocity systems. This ensures that the feedforward compensation command accurately covers the actual needs of the controlled object within an extremely wide flow adjustment range, avoiding reverse regulation and system oscillations caused by time-domain mismatch.

[0017] 3. Based on frequency domain decoupling, source-load response matching, and inherent stability construction, the total heat load demand is decoupled in the frequency domain into low-frequency base quantity and high-frequency increment, which are mapped to the inertial base load group and the agile adjustment group, respectively. This solves the response mismatch and chasing effect in large inertial systems. Moving average filtering is used to extract the trend term as the base load command, shielding the impact of high-frequency fluctuations on large inertial equipment and maintaining efficient and stable operation. The agile group is used to perform transient compensation only for high-frequency residuals, giving full play to the advantages of rapid adjustment. The frequency division governance control architecture avoids the phenomenon of over-adjustment caused by forcibly requiring large lagging equipment to respond quickly to disturbances. Without adding additional damping links, the inherent stability and adjustment quality of the system are improved through the structured design of control logic. Attached Figure Description

[0018] Figure 1 is a schematic diagram of the frequency division decoupling control logic and multi-dimensional feedback closed loop principle of the present invention; Figure 2 is a graph of the total heat load demand and the response curves of the high and low frequency components after decoupling under the variable operating conditions of the present invention; Figure 3 is a schematic diagram of the system hardware layer deployment architecture and network topology based on cloud-edge collaboration of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be described in detail below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] The AI-based heat load prediction and multi-heat source optimization scheduling system provided by this invention mainly includes: a source characteristic storage unit, a load trend prediction unit, a control command generation unit, a collaborative scheduling unit, and a feedback correction unit. These units work together to form a composite control system with trend feedforward, frequency division decoupling, and targeted feedback. The source characteristic storage unit corresponds to the non-volatile storage area within the system controller, used to store the physical characteristic parameters of each heat source device in the heating network, establish a mathematical model of heterogeneous actuators based on the response time constant, and provide logical criteria for frequency division scheduling. Key parameters include the response time constant obtained through offline step response testing. The system includes time constants, maximum allowable ramp rates, and operating energy efficiency cost parameters for each device. During system initialization, based on a preset baseline value (e.g., 15 minutes), the response time constant is compared. Heat source devices with values ​​exceeding this baseline value, such as ground source heat pump units, industrial waste heat recovery devices, or biomass boilers, are classified into an inertial base load group. Heat source devices with values ​​below this baseline value, such as gas-fired boilers or electrode boilers, are classified into an agile adjustment group. This establishes a physical grouping basis for subsequent control strategy distribution. A load trend prediction unit is used to predict heat load change trends to address the thermal inertia and transmission lag of the heating network. This unit serves as the feedforward compensation core of the control system and employs a time series prediction model. The system employs either a Long Short-Term Memory (LSTM) network model or a gated recurrent unit (GRU) network model. By extracting nonlinear features from historical operating parameters, it constructs a demand evolution model for the controlled object, providing advanced control variables for coordinated scheduling. In deployment, this model runs on edge computing nodes adjacent to the heating area, acquiring historical operating data and environmental parameters of the heating area via an industrial fieldbus. The historical operating data includes outdoor temperature, supply water temperature, and return water temperature sequences. This unit calculates and outputs the slope of heat load change within the future pipeline transmission lag period, outputting a trend value such as +0.1 MW / h. This slope serves as the basis for feedforward control. A control command generation unit is used for... The total heat load demand, which includes high-frequency disturbances, is decoupled by extracting real-time features of the load signal in the frequency domain, achieving a dual decomposition of the control target in both the time and frequency domains. The system receives the real-time total heat load demand signal and, based on the heat load change slope output by the load trend prediction unit, processes the total heat load demand using a low-pass filtering algorithm, such as a digital moving average filter or a Butterworth low-pass filter. The filtering process decomposes the total demand into two components: the output of the filtering algorithm is the low-frequency base demand, representing the slow trend of load change; the residual value after subtracting the low-frequency base demand from the real-time total heat load demand is the high-frequency incremental demand, representing rapid fluctuations and disturbances in the load.

[0021] The cutoff frequency of the low-pass filter algorithm is dynamically adjustable. It monitors the remaining adjustable capacity of the agile control group in real time. When the output power of the agile control group approaches its upper limit, causing the remaining adjustable capacity to fall below a preset safety value (e.g., below 10%), the cutoff frequency is actively lowered. This allows more mid-to-high frequency disturbance components to be allocated to the low-frequency base load demand, increasing the load proportion allocated to the inertial base load group and preventing the agile control group from losing its control capability due to overload. The collaborative scheduling unit executes the frequency division mapping mechanism, mapping demand components of different frequencies to heat source groups with corresponding characteristics. The low-frequency base load demand output by the control command generation unit is converted into a first control command for the inertial base load group. This command changes gradually, with the rate of change limited to within the maximum allowable ramp rate of the inertial base load group, ensuring the base load heat source can smoothly change. The system operates stably within its high-efficiency range. This unit converts high-frequency incremental demand into second control commands for the agile control group, utilizing the rapid response characteristics of the agile heat source to offset instantaneous fluctuations in the pipeline network. A feedback correction unit provides closed-loop feedback correction. This unit does not measure absolute temperature but instead collects the rate of change of the total return water temperature in the heating network in real time. By monitoring the rate of change of the controlled variable in real time, an error feedback loop is constructed to ensure the system maintains dynamic stability when subjected to high-frequency random disturbances. To improve signal quality, this rate of change is obtained by differential calculation of the return water temperature measurements from multiple consecutive sampling periods, followed by smoothing and filtering, such as Gaussian filtering. The collaborative scheduling unit further executes targeted compensation logic based on this rate of change, pre-setting a stable dead zone range. The width of this dead zone range is as follows: The minimum adjustable power step size is determined based on the agile control group to prevent oscillations near zero. When the rate of change of the total return water temperature of the heating network exceeds the stable dead zone, it indicates a supply-demand imbalance. The coordinated scheduling unit immediately generates a correction compensation amount and executes targeted operations: the correction compensation amount is added only to the second control command to adjust the output power of the agile control group, while maintaining the first control command unchanged, ensuring that the high-frequency correction action is undertaken by the agile heat source, protecting the stable operation of the base load heat source. The load trend prediction unit is also used to perform feedforward time window adaptive adjustment. This scheme adds a flow-time mapping unit, which is used to collect the total circulating flow data of the heating network in real time and call the preset flow-time nonlinear mapping relationship table established through offline calibration or fluid simulation. This unit looks up the current flow data and sets the current flow data as follows: The equivalent transmission lag time, such as 35 minutes, is found and converted. The load trend prediction unit dynamically sets this equivalent transmission lag time as the effective prediction step size of the time series prediction model to ensure that the prediction time span of the heat load change slope is synchronized with the current hydraulic transport speed of the pipeline network in real time.

[0022] In another embodiment, to address the issue of long-term occupation of high-cost heat sources due to load characteristic evolution, the collaborative scheduling unit is also used to execute energy reference dynamic migration logic. This scheme adds a reference migration arbitration unit, which calculates the integral average of the second control command within a preset sliding time window longer than the pipeline transmission lag period, such as 60 minutes. When the duration of this integral average exceeds a preset dwell threshold, it is used to distinguish between transient disturbances and steady-state load drift, determining that the load center has undergone substantial drift. At this time, the collaborative scheduling unit generates a reference migration compensation amount whose rate of change is limited by the maximum ramp rate of the inertial base load group. This unit then adds the reference migration compensation amount to the first control command and simultaneously subtracts it from the second control command. The control variable is equal to the baseline migration compensation value. This process continues until the integral mean returns to the preset balance dead zone range, transferring the steady-state load borne by the agile heat source to the base load heat source. To address the performance degradation of the physical actuator due to aging, scaling, and other factors, which leads to control model mismatch, the system also includes an efficiency adaptive correction unit. This unit is used to compensate for the performance degradation of the physical actuator online. It calculates the steady-state residual between the theoretical expected response and the actual observed value to achieve online self-correction of the control system gain coefficient, ensuring that the control law always remains consistent with the real-time physical characteristics of the actuator. The working logic is based on the current second control command and the preset initial equipment gain coefficient, such as... =1.0, calculate the theoretical expected response of the agile adjustment group, such as +0.5 theoretically. / min The theoretically expected response is compared with the response provided by the feedback correction unit. The actual measured value, such as only +0.4. / min, perform difference calculations to obtain the steady-state residual value. When the steady-state residual value continuously exceeds the preset fault determination threshold, the unit determines that the equipment performance has degraded; based on the steady-state residual value, the unit updates the initial equipment gain coefficient in reverse using the following logic: ,in, This is the updated gain coefficient. The gain coefficient currently in use. The preset correction step size factor is, for example, 0.05. For steady-state residual values, The system uses the updated gain coefficient as a reference value for the rated response of the agile adjustment group. The second control command is weighted and modified to increase the output force.

[0023] Example 1: In a centralized heating system coupling a ground source heat pump unit and a gas boiler, the system is operating under variable flow conditions to cope with peak daytime demand. At this time, the total circulating flow of the pipeline network is high, and the corresponding equivalent transmission lag time is short. When a sudden cold wave causes the outdoor ambient temperature to drop rapidly by 5°C within 1 hour... At this time, the system control logic responds as follows: the load trend prediction unit calculates the slope of heat load change within the future pipeline transmission lag period based on real-time collected environmental parameters, determining it to be a high-intensity positive growth trend; the flow-time mapping unit, based on the current high circulation flow data, calls a preset flow lag time nonlinear mapping table, automatically shrinking the effective prediction step size of the time series prediction model to a shorter time value to match the current accelerated hydraulic transport speed; ensuring that the steep slope of change output by the load trend prediction unit is synchronized in time phase with the physical moment when the heat demand wavefront arrives at the end; the control command generation unit receives the steep slope of change and the real-time increasing total heat load demand, and uses low-pass filtering to calculate... The method decouples this demand into two parts: a rapidly rising low-frequency base demand, corresponding to a continuous downward trend in temperature, and a high-frequency incremental demand with a large amplitude in the initial stage of change, corresponding to the impact of a sudden drop in temperature. The coordinated scheduling unit then maps the low-frequency base demand to the first control command for the inertial base load group, and begins to slowly increase the output at its maximum allowable ramp rate. The high-frequency incremental demand is mapped to the second control command for the agile adjustment group, which enables the gas boiler to start and stop quickly or increase its output, instantly compensating for most of the heat gap. When the sudden cold wave turns into a sustained low temperature, the high-frequency disturbance evolves into a steady-state load, and the second control command of the agile adjustment group is maintained at a high output state for more than 30 minutes.

[0024] The benchmark migration arbitration unit calculates the integral mean of the second control command within a preset sliding time window and determines that it has exceeded a preset dwell threshold, indicating that the load center has experienced steady-state drift. The system then initiates the energy benchmark dynamic migration logic, and the coordinating scheduling unit generates a benchmark migration compensation amount with an extremely low rate of change. This compensation amount is added to the first control command, and an equal control amount is simultaneously subtracted from the second control command. This process smoothly transfers the steady-state load originally borne by the high-cost gas boiler to the low-cost ground source heat pump unit until the integral mean returns to the equilibrium dead zone. Within the zone; throughout the entire operating condition switching and load migration process, the feedback correction unit continuously monitors the rate of change of the total return water temperature of the heating network. Any supply and demand imbalance caused by prediction deviation or actuator nonlinearity is reflected as the rate of change exceeding the stable dead zone range, thereby generating a correction compensation amount and superimposing it only on the second control command. The high-speed adjustment capability of the gas boiler is used to eliminate high-frequency thermal hysteresis fluctuations. The synergistic effect of frequency division decoupling, trend feedforward and targeted feedback enables the system to maintain a stable terminal heating temperature when dealing with severe external disturbances, and maximizes the operating proportion of the low-cost inertial base load group.

[0025] Example 2: On a simulation test platform for evaluating the control strategy, a heating system model was built, including a ground source heat pump unit as the inertial base load group and a gas boiler as the agile regulation group; a standard 24-hour heat load disturbance curve was set, which included the rapid step cooling simulating cold wave disturbance and the high-frequency random fluctuation of noise simulating actual operating conditions. The physical parameter setting was a transmission lag time of 40 minutes for the heating network; four sets of tests were set up: control group 1 adopted a traditional PID feedback control strategy based on return water temperature measurement; control group 2 adopted the frequency division decoupling and targeted feedback of the present invention, with the load deactivated. Trend prediction unit; control group 3 adopts the frequency division decoupling and trend feedforward of the present invention, and disables the feedback correction unit; the sample group of the present invention adopts the complete technical solution mentioned above, that is, simultaneously enabling frequency division decoupling, trend feedforward and targeted feedback; under the same standard heat load disturbance curve, the four groups of tests run continuously for 24 hours, and record the maximum deviation of the water supply temperature at the end of the heating network when dealing with rapid step cooling disturbance, which is used to characterize the control stability and the percentage of the cumulative operating heat supply of the agile adjustment group to the total heat supply during the entire operating cycle, which is used to characterize the operating economy. The test data are shown in Table 1.

[0026] Table 1: Comparison of Stability and Economy under Different Control Strategies

[0027] Table 1 shows that the traditional PID strategy in control group 1 exhibits temperature overshoot and oscillation, with the agile control group accounting for 41.5% of operations; in control group 2, the scheme that disables the load trend prediction unit reduces the temperature deviation to 1.8. The agile adjustment group still accounted for 33.2% of the operation; in control group 3, the scheme that discontinued the feedback correction unit, the temperature deviation and oscillation frequency were both higher than those in control group 2; the maximum deviation of the terminal water supply temperature in the sample group of this invention was 0.7. The pipeline temperature oscillations were minimized, the proportion of agile adjustment groups in operation dropped to 18.6%, and the combined effect of load trend feedforward and targeted feedback maintained the stability of heating temperature and increased the proportion of low-cost base load heat sources in operation.

[0028] Example 3: This example, in conjunction with Figures 1 to 3, describes an AI-based heat load prediction and multi-heat source optimization scheduling system. As shown in Figure 1, the system architecture starts with the edge computing node data source, collects historical operating data, environmental parameters, and real-time flow, and transmits them downstream. The flow-time mapping unit is responsible for performing nonlinear mapping of flow rate and lag time and adaptively adjusting the prediction step size, outputting the equivalent transmission lag time to the load trend prediction unit. The load trend prediction unit calculates the heat load change slope based on environmental and load data using a time series prediction model, and inputs it into the control command generation unit. Combining the device grouping and parameters provided by the source characteristic storage unit, low-pass filtering is used... The algorithm decoupling and dynamically adjustable cutoff frequency decompose the demand into decoupled low-frequency and high-frequency demands. Based on the frequency division mapping mechanism and energy benchmark dynamic migration logic, the collaborative scheduling unit generates a first control command (low-frequency base quantity) to be delivered to the inertial base load group, which includes ground source heat pumps and waste heat recovery. Simultaneously, it generates a second control command (high-frequency increment) to be delivered to the agile adjustment group, which includes gas-fired boilers and electrode boilers. The system constructs a dual closed-loop feedback: first, the feedback correction unit collects the total return water temperature of the heating network and its rate of change to generate targeted correction compensation amounts that act on the collaborative scheduling unit; second, the efficiency adaptive correction unit calculates the steady-state residual based on the total return water temperature of the heating network and updates the initial equipment gain coefficient. This updates the gain coefficient. This feedback is sent to the control loop.

[0029] As shown in Figure 2, the horizontal axis represents time in minutes, ranging from 0 to 55 minutes. The vertical axis represents heat load demand in megawatts (MW). The figure contains three lines: the solid line represents the total heat load demand, which initially rises and then steadily declines over time; the dashed line represents the low-frequency baseline demand after filtering, which is smooth and lags behind the total demand change; and the dotted line represents the high-frequency incremental demand, which initially rises sharply with the total load from 0 to 25 minutes and shows a positive value, then rapidly falls back to near zero or even negative values ​​after the total load stabilizes. As shown in Figure 3, the system's hardware deployment architecture is divided into three layers from top to bottom. The top layer is a remote cloud training platform responsible for long-cycle model reconstruction and distributing new model parameters. The middle layer... As the intelligent decision-making and control hub, it includes edge computing nodes as the AI ​​brain and a multi-heat source collaborative controller PLC. The edge computing nodes are equipped with load trend prediction units and flow time mapping units to output prediction slopes and feedforward signals. The PLC is equipped with frequency division decoupling and instruction generation modules as well as feedback correction and adaptive modules. It connects to the underlying devices through an industrial fieldbus as a high-speed data and instruction path. The underlying devices are divided into three areas: the left wing is the inertial base load execution area, which includes ground source heat pump units and industrial waste heat recovery devices to respond to low-frequency base load instructions; the right wing is the agile adjustment execution area, which includes gas boilers and electrode boilers to respond to high-frequency incremental instructions; and the middle is the network status perception area, which is equipped with a total return water temperature sensor and a pipeline total circulation flow meter to upload feedback data in real time.

[0030] Example 4: The calibration procedure for the stable dead zone range required for the operation of the feedback correction unit is as follows: During the offline commissioning phase of the system, lock the inertial base load group, start the agile adjustment group and make it work on the stable base load; apply a step command with the minimum control step size to the agile adjustment group to increase the output power to the minimum adjustment unit; maintain this command for 5 minutes, and simultaneously collect the total return water temperature of the heating network; repeat this step test at least 20 times; for each test, calculate the rate of change of return water temperature within 5 minutes after the step. The mean; obtained from these 20 trials Mean statistical analysis, calculation of standard deviation And set the stability dead zone range to To ensure that the control logic does not respond to normal temperature fluctuations caused by the minimum adjustment action of the agile adjustment group; the calibration procedure for the core parameters required for the coordinated scheduling unit to execute the energy reference dynamic migration logic is as follows: By analyzing historical data, the duration of typical transient load disturbances is statistically analyzed, and the 95th percentile is taken as the transient disturbance period. ,like =30 minutes; Set the duration of the sliding time window to... That is, 60 minutes, to ensure that the window length is sufficient to cover the entire transient disturbance and avoid misjudging it as a steady-state drift; set the dwell threshold to That is, 30 minutes. When the integral average of the second control command deviates from zero for more than 30 minutes, the system determines that steady-state load drift has occurred; the balance dead zone range is set to a smaller integral average range, such as the corresponding agile adjustment group. The rated power output is used to stop the continued accumulation of the reference migration compensation amount after the migration is completed.

[0031] The internal algorithm of the performance adaptive correction unit is implemented as follows: This unit calculates the theoretically expected response based on a simplified thermodynamic model. By comparing the expected output of the control model with the actual feedback from the controlled object in real time, the model executes parameter reverse identification and update logic. ,in, It is the theoretically expected rate of change of return water temperature. The second control command is sent to the power value of the agile adjustment group. This is the preset initial device gain coefficient, with an initial value of 1.0. It is the specific heat capacity of water. This is the current real-time total circulating flow of the pipeline network; this unit also receives data from the feedback correction unit. Actual measured value ; Calculate the original residual sequence To extract the steady-state residual values, a first-order low-pass filter with the same time constant as the baseline migration window, such as 60 minutes, is used to process the original residual sequence. ,in, These are the filter coefficients. The extracted steady-state residual value; when this If the failure rate consistently exceeds the preset fault threshold, such as 5% of the rated response of the corresponding agility adjustment group, the gain coefficient update logic will be activated. and utilize the updated replace Used for subsequent calculations of the theoretical expected response.

[0032] Example 5: During the initial deployment of the system, an offline calibration procedure needs to be executed to determine the parameters of the flow-time mapping unit. This procedure stabilizes the total circulating flow rate of the pipeline network at a preset low value, such as 30% of the rated flow rate. A brief heat pulse is generated at the heat source inlet, and a temperature sensor is used to monitor the return water temperature at the farthest end of the pipeline network. The propagation time from the issuance of the heat pulse to the start of the temperature response at the farthest end is recorded. This time is calibrated as the equivalent transmission lag time under the current flow rate. This test process is repeated at multiple different flow rate setpoints, such as from 30% to 100% in increments of 5%. All measured total circulating flow rate data points and corresponding equivalent transmission lag time data points are used to fit a nonlinear curve and solidify it into a preset flow lag time nonlinear mapping. Relationship table; The initial model training procedure for the load trend prediction unit is as follows: In the initial stage of system deployment, a data accumulation period is set, taking 14 days as an example. During this period, the control system operates using the basic control strategy and records all relevant historical operating data and environmental parameters at a sampling frequency of no less than once every 5 minutes. The data constitutes the initial training set. After the end of this period, the initial training set is used to train the time series prediction model offline. The goal of the training is to make the root mean square error of the model's prediction of the slope of the heat load change during the future pipeline transmission lag period lower than the preset engineering threshold, such as lower than 2% of the total load fluctuation. When the model converges and meets the target, the trained model is loaded to the edge computing node and enters the online prediction operation state.

[0033] Example 6: Low-pass filtering algorithm for the control command generation unit, cutoff frequency Based on the remaining adjustable capacity of the agile adjustment group Dynamic adjustment Defined as 1 minus the ratio of the current output power of the agile adjustment group to its maximum power; The adjustment logic follows a preset non-linear mapping relationship, which in the specific implementation is as follows: ,in Based on the cutoff frequency, The maximum cutoff frequency, For an exponential factor greater than 1, taking 2 as an example, this setting makes the remaining capacity... When sufficient, cutoff frequency Maintain at Nearby, more high-frequency components are allowed to enter the agile group, and when When it decreases, Rapidly decreasing in a non-linear manner More load is forcibly allocated to the inertial base load group; in the simulation, the actual output gain of the agile adjustment group is artificially set to its initial device gain coefficient. 90% of the performance was simulated, with a 10% performance degradation in the physical actuator; a constant high-frequency incremental demand was applied to the system; under control conditions with the performance adaptive correction unit disabled, the theoretically expected response was observed to be... The actual measured values ​​produce a steady-state residual value corresponding to a continuous output deviation of approximately -10%. Under the same operating conditions, the performance adaptive correction unit is activated, and based on the steady-state residual value, the aforementioned method is applied. The update logic adjusts the internal gain coefficient within a correction cycle of approximately 30 minutes. Gradually increasing to approximately 1.11, as The update automatically amplifies the second control command, increasing the actual output power of the agile adjustment group, and the steady-state residual value. The load trend prediction unit also includes an online model performance monitoring and periodic reconstruction mechanism, which calculates the root mean square error of the time series prediction model over a long period, such as 24 hours, in real time. When this error exceeds the engineering threshold of the initial training for three consecutive days, i.e., 2% of the total load fluctuation, the system determines that the model has mismatched. At this time, the system automatically uses the historical running data of the most recent 30 days as a new training set and starts a low-priority model retraining task in the background. After the new model training converges, the system deploys it in shadow mode to run in parallel with the current online model without outputting control commands. When the prediction error of the new model in shadow mode is lower than that of the current online model for 24 consecutive hours, the system switches the new model to the online running model. This procedure ensures the adaptability of the model and does not affect the stability of online operation.

[0034] In the low-pass filtering algorithm cutoff frequency dynamic adjustment logic used in the control command generation unit, the offline calibration procedure for key parameters includes: the basic cutoff frequency. To determine this, in the system simulation environment, the agile adjustment group is locked at its maximum output power. At this point, a step load command is applied to the inertial base load group, and the time required for the output power to reach 95% of the set value is collected. , the fundamental cutoff frequency Set as The corresponding values ​​ensure that the filtering frequency does not exceed the physical response limit of the base load group when the agile group is fully loaded; maximum cutoff frequency. The determination was made using the same method, under the condition of inertial base load group locking, by measuring the response time of the agile adjustment group. The maximum cutoff frequency Set as Corresponding value; exponential factor The factor was initially set to 2, and the remaining adjustable capacity of the agile adjustment group was simulated in the system simulation platform. As the operating rate decreases from 100% to 10%, the rate of change in the total return water temperature in the pipe network gradually increases. value, until the rate of change is When the percentage drops below 20%, the convergence rapidly reaches the stable dead zone. The value is a determined optimized value; the preset benchmark value for heat source division is determined for the source characteristic storage unit. The engineering calibration procedure includes: analyzing historical load data of the entire heating season, extracting all transient disturbance events with a duration shorter than the energy benchmark dynamic migration residence threshold, such as 30 minutes, and calculating the 95th percentile value of the duration of these transient disturbance events. and this A value of 15 minutes is set as a preset baseline value to match the physical response capability of the agile adjustment group with the instantaneous disturbance to be processed on a time scale. The fault judgment threshold calibration procedure of the efficiency adaptive correction unit is as follows: under the condition that the agile adjustment group is confirmed to have no physical fault and is operating at 50% rated power, the original residual sequence between the theoretical expected response and the actual measured value is continuously collected for at least 24 hours, and the output value of the sequence after steady-state residual extraction logic processing is calculated. Standard deviation Set the fault determination threshold to .

[0035] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0036] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. An AI-based heat load prediction and multi-heat source optimization scheduling system, characterized in that, include: The source characteristic storage unit is used to store the response time constant of each heat source device in the heating network, and divides the heat source devices into an inertial base load group and an agile adjustment group according to the response time constant. The inertial base load group has a response time constant greater than the preset reference value, and the agile adjustment group has a response time constant less than the preset reference value. The load trend prediction unit calculates the slope of heat load change in the heating area during the future pipeline transmission lag period based on historical operating data and environmental parameters of the heating area using a time series prediction model. The control command generation unit decouples the real-time total heat load demand into low-frequency base demand and high-frequency incremental demand using a low-pass filtering algorithm based on the heat load change slope. The collaborative scheduling unit establishes a frequency-division mapping mechanism, which maps the low-frequency base demand to a first control command for the inertial base load group and maps the high-frequency incremental demand to a second control command for the agile regulation group. The rate of change of the first control command is limited to the allowable ramp rate of the inertial base load group. The feedback correction unit collects the rate of change of the total return water temperature of the heating network in real time. The collaborative scheduling unit also executes targeted compensation logic. When the rate of change of the total return water temperature of the heating network exceeds the preset stable dead zone range, a correction compensation amount is generated and superimposed only on the second control command to adjust the output power of the agile regulation group while keeping the first control command unchanged, thereby using the agile regulation group to eliminate high-frequency thermal lag fluctuations.

2. The AI-based heat load prediction and multi-heat source optimization scheduling system according to claim 1, characterized in that, The load trend prediction unit is also used to perform feedforward time window adaptive adjustment. The adjustment includes: the flow-time mapping unit is used to collect the total circulation flow data of the heating network in real time, and call the preset flow and lag time nonlinear mapping relationship table to find the equivalent transmission lag time corresponding to the total circulation flow data; the load trend prediction unit sets the equivalent transmission lag time as the effective prediction step size of the time series prediction model, so that the prediction time span of the heat load change slope is kept synchronized with the current hydraulic transport speed of the network in real time.

3. The AI-based heat load prediction and multi-heat source optimization scheduling system according to claim 1, characterized in that, The coordinated scheduling unit is also used to execute the energy reference dynamic migration logic, which includes: the reference migration arbitration unit is used to calculate the integral average of the second control command within a preset sliding time window; when the integral average continuously exceeds the preset dwell threshold, the coordinated scheduling unit generates a reference migration compensation amount, the rate of change of the reference migration compensation amount is limited by the maximum allowable ramp rate of the inertial base load group; the coordinated scheduling unit adds the reference migration compensation amount to the first control command, and simultaneously subtracts a control amount equal to the value of the reference migration compensation amount from the second control command, until the integral average returns to the preset balance dead zone range.

4. The AI-based heat load prediction and multi-heat source optimization scheduling system according to claim 3, characterized in that, The sliding time window executed by the benchmark migration arbitration unit is longer than the pipeline transmission lag period, and the dwell threshold is set as the energy judgment boundary to distinguish between transient disturbances and steady-state load drift.

5. The AI-based heat load prediction and multi-heat source optimization scheduling system according to claim 1, characterized in that, The system also includes an adaptive performance correction unit to compensate for the performance degradation of the physical actuator. The adaptive performance correction unit is configured to calculate the theoretical expected response of the agile adjustment group based on the second control command and the preset initial equipment gain coefficient. The steady-state residual value is obtained by differential calculation between the theoretical expected response and the actual measured value of the rate of change of the total return water temperature of the heating network. When the steady-state residual value exceeds the preset fault judgment threshold, the initial equipment gain coefficient is updated in reverse according to the steady-state residual value, and the second control command is weighted and corrected using the updated gain coefficient.

6. The AI-based heat load prediction and multi-heat source optimization scheduling system according to claim 5, characterized in that, The performance adaptive correction unit updates the initial device gain coefficient using the following logic: ,in, This is the updated gain coefficient. The gain coefficient currently in use. The preset correction step size factor, For steady-state residual values, This is the reference value for the rated response of the agile adjustment group.

7. The AI-based heat load prediction and multi-heat source optimization scheduling system according to claim 1, characterized in that, The inertial base load group includes ground source heat pump units, industrial waste heat recovery devices, or biomass boilers; the agile adjustment group includes gas-fired boilers or electrode boilers.

8. The AI-based heat load prediction and multi-heat source optimization scheduling system according to claim 1, characterized in that, The load trend prediction unit uses a long short-term memory network model or a gated cyclic unit network model, which is deployed in edge computing nodes near the heating area. The edge computing nodes directly obtain historical operating data and environmental parameters through the industrial fieldbus.

9. The AI-based heat load prediction and multi-heat source optimization scheduling system according to claim 1, characterized in that, The cutoff frequency of the low-pass filtering algorithm used by the control command generation unit is dynamically adjustable. The control command generation unit adjusts the cutoff frequency in real time according to the remaining adjustable capacity of the agile adjustment group. When the remaining adjustable capacity is lower than the preset safety value, the cutoff frequency is reduced to increase the proportion of low-frequency base quantity demand allocated to the first control command.

10. The AI-based heat load prediction and multi-heat source optimization scheduling system according to claim 1, characterized in that, The rate of change of the total return water temperature of the heating network collected by the feedback correction unit is obtained by performing differential calculation on the return water temperature measurement values ​​of multiple consecutive sampling cycles and then smoothing and filtering them. The preset stable dead zone range is determined based on the minimum adjustable power step size of the agile adjustment group.

Images