Multi-energy complementary heat supply intelligent regulation method and system of solar heat supply system

Abstract

The application provides a multi-energy complementary heating intelligent regulation method and system of a solar heating system, the method comprises the following steps: collecting photovoltaic system and heating pipe network multi-source operation data to construct a time sequence input matrix, segmenting the data state according to irradiance, cloud cover and heat load change rate, inputting each segment into a long short-term memory network, calculating a forgetting gate basis value, adjusting the forgetting coefficient when the cloud cover suddenly changes by combining the continuous prediction residual error, obtaining the state retention amount through weighted alignment, generating a memory compression factor according to the state reversal index, obtaining the state write amount by constraint, outputting the solar output sequence and fluctuation index after updating the unit state, constructing a target function based on the sequence, fluctuation index, heat load prediction, heat storage state and electricity price, solving to obtain the time period distribution of the photovoltaic heat pump, gas boiler, heat storage and water supply temperature, adjusting the threshold value by combining the prediction credibility, and generating a multi-energy complementary heating regulation instruction.

Description

Technical Field

[0001] This application belongs to the field of regulation and control, and in particular relates to a multi-energy complementary heating intelligent regulation and control method and system for a solar heating system. Background Technology

[0002] With the accelerated transformation of the energy structure, the penetration rate of renewable energy in the heating sector is continuously increasing. Multi-energy complementary heating systems, consisting of photovoltaic systems, heat pumps, gas boilers, and thermal storage devices, have become a key path to improve energy utilization efficiency. Solar power output is greatly affected by meteorological conditions such as irradiance and cloud cover, exhibiting strong randomness and volatility. This instability at the energy supply end makes it impossible to control the energy input on the source side of the heating network, easily leading to severe supply-demand mismatch between the heating source and the changing terminal heat load demand. Without accurate output forecasting and multi-source coordinated scheduling, there is often a serious problem of solar curtailment, or excessive reliance on high-cost gas boilers for peak shaving during periods of heating shortage, undermining the system's initial energy-saving goals and severely reducing the overall network's economic efficiency and heating reliability. Existing technologies often utilize deep learning models such as Long Short-Term Memory (LSTM) networks to process multi-source time-series data for solar power output forecasting, and combine the forecast data with heat load and electricity prices to formulate heating control schemes. LSTM networks, through internal control mechanisms, can extract long-term dependencies between meteorological and load sequences, assisting in optimizing multi-energy complementary allocation.

[0003] However, the forget gate update mechanism of conventional LSTM is relatively rigid. When cloud cover changes abruptly or irradiance changes drastically, it cannot adjust based on the segmentation of prediction residuals and input features. It cannot accurately represent the impact of irradiance slope, cumulative temperature difference, and load gradient on the hidden state, which easily leads to prediction lag and failure to detect state reversal. Existing control methods usually rely only on a single accurate prediction value and fail to combine output sequence fluctuation indicators to evaluate prediction reliability. As a result, when operating conditions change abruptly or thermal storage is limited, it is impossible to adjust the system's energy absorption and load gap threshold. The generated allocation and control commands cannot cope with multi-source uncertainties and cannot achieve highly stable multi-energy complementary collaborative optimization. Summary of the Invention

[0004] To address the problems of rigid and inaccurate forecasting in existing technologies, and the fact that regulation does not take into account output fluctuations and forecast reliability, making it difficult to cope with energy uncertainties, resulting in a mismatch between heating supply and demand, and low system economy and reliability.

[0005] In a first aspect, the present invention provides a multi-energy complementary heating intelligent control method for a solar heating system, comprising: A time-series input matrix is ​​constructed by collecting multi-source operation data of photovoltaic system and heating network. The time-series input matrix is ​​segmented into states according to irradiance, cloud cover and heat load change rate, and the irradiance slope, temperature difference accumulation, load gradient and state reversal index of each segment are calculated. Each segment is input into the Long Short-Term Memory Network to calculate the base value of the forget gate based on the current input and the previous hidden state. When cloud cover changes abruptly, the base value of the forget gate is adjusted based on the continuous prediction residual characteristics to obtain the updated forgetting coefficient. The updated forgetting coefficient and the previous unit state are jointly weighted and aligned in dimension according to the dimensionless irradiance slope, temperature difference accumulation and load gradient to obtain the state retention amount. The memory compression factor is generated according to the state reversal index. The state writing amount is obtained by constraining the input gate output and candidate states. The state retention amount and the writing amount are superimposed to update the unit state and generate the solar power output sequence and fluctuation index. An objective function is constructed based on the solar power output sequence, fluctuation index, heat load forecast, thermal storage status, and electricity price sequence. The objective function and constraint equation set are solved to obtain the time-period allocation of photovoltaic-driven heat pumps, gas boilers, thermal storage charging and discharging, and water supply temperature. The prediction reliability of the solar power output sequence is calculated. The heat pump absorption and load gap thresholds are adjusted according to the prediction reliability. Multi-energy complementary heating control instructions are generated by combining the time-period allocation.

[0006] Furthermore, the construction of a time-series input matrix from the multi-source operational data of the photovoltaic system and heating network includes: Historical discrete time series of photovoltaic array irradiance, module temperature, ambient temperature, cloud cover, heating network supply and return water temperature, heat load, electricity price, and state of charge of thermal storage units are acquired using data acquisition sensors. For time series with inconsistent frequencies and misalignments, the arithmetic mean interpolation of adjacent time points is used to align the data nodes and integrate the time series into a unified sampling frequency with a consistent time interval. The aligned historical data points are then stacked and merged as independent input vectors in the time domain to generate a two-dimensional time series input matrix that expresses the complete temporal development characteristics of multi-source data.

[0007] Furthermore, the step of segmenting the time-series input matrix into states based on irradiance, cloud cover, and heat load change rate, and calculating the irradiance slope, cumulative temperature difference, load gradient, and state reversal index for each segment, includes: The time-series input matrix is ​​processed using a time sliding window to calculate the difference between parameters corresponding to adjacent time nodes in the time dimension, resulting in the irradiance change rate, cloud cover change rate, and heat load change rate. Waveform regions where the absolute value of the change rate continuously exceeds a set segment constant are extracted as unstable abrupt change segments. The remaining main data is truncated at a set fixed interval to form relatively static segments, thus constituting independent state segments. The irradiance slope is obtained by dividing the irradiance difference at the start and end positions of the state segment by the time span of the segment. The temperature difference accumulation is obtained by summing the absolute values ​​of the component temperature and ambient temperature differences at each moment within the state segment. The load gradient is obtained by dividing the heat power difference at the start and end positions of the state segment by the time span of the segment. The state reversal index is obtained by counting the cumulative number of sign reversals of the time-domain waveform trends of parameters within the state segment, from rising to falling or from falling to rising.

[0008] Further, the step of inputting each segment into the Long Short-Term Memory network to calculate the base value of the forgetting gate based on the current input and the previous hidden state, adjusting the base value of the forgetting gate based on the continuous prediction residual characteristics when cloud cover changes abruptly to obtain the updated forgetting coefficient, and jointly weighting the updated forgetting coefficient and the previous unit state according to the dimensionless irradiance slope, cumulative temperature difference, and load gradient, and aligning the dimensions to obtain the state retention amount, includes: The forget gate base value is calculated by inputting the operation matrix. The forget gate base value is equal to the sum of the product of the Sigmoid activation function applied to the correlation matrix and the concatenation result of the current input and the previous hidden state, plus the bias. When the change in cloud cover is greater than the limit threshold, the continuous prediction bias is used to intervene and obtain the updated forget coefficient. The updated forget coefficient is equal to the forget gate base value minus the product of the adjustment step difference limit rate and the difference of the normalized residual, and the value is truncated to a boundary between 0 and 1. If the change in cloud cover is not greater than the limit threshold, the updated forget coefficient is equal to the forget gate base value. The state retention is equal to the fusion vector obtained by adding the first, second, and third mapping vectors element by element. This fusion vector is then calculated by performing Hadamard products on the updated forgetting coefficient vector and the previous unit state vector. The first mapping vector is the product of the first training coefficient vector and the dimensionless irradiance slope scalar. The second mapping vector is the product of the second training coefficient vector and the temperature difference accumulation scalar. The third mapping vector is the product of the third training coefficient vector and the load gradient scalar. The first, second, and third training coefficient vectors are all learnable parameter vectors with the same dimension as the unit state.

[0009] Further, the step of generating a memory compression factor based on the state reversal exponent, constraining the input gate output and candidate states to obtain the state write amount, superimposing the state retention amount and the write amount to update the unit state, and generating a solar power output sequence and fluctuation index includes: A memory compression factor is generated using the natural exponential function. The memory compression factor is equal to the product of the negative decay factor constraint term of the natural exponential base and the state reversal exponent raised to the power of the power of the product. The state write amount is calculated by scaling the new variable adsorption vector using the memory compression factor. The state write amount is equal to the product of the memory compression factor and the Hadamard product of the input gate output and the candidate state. The updated unit state is then superimposed, and the unit state is equal to the sum of the state retention amount and the state write amount. Based on the updated unit state, the future time-domain solar power output sequence is generated by projecting the prediction delay execution sequence through a fully connected layer. A fluctuation index is calculated using a quadratic metric formula. The fluctuation index is equal to the square root of the sum of the squares of the differences between the power at each time step and the average power of the solar power output sequence over the total time step span.

[0010] Furthermore, the step of constructing an objective function based on solar power output sequence, fluctuation index, heat load forecast, thermal storage status, and electricity price sequence, and solving the objective function and constraint equation set to obtain the time-period allocation of photovoltaic-driven heat pumps, gas boilers, thermal storage charging and discharging, and water supply temperature includes: An objective function is constructed with total cost minimization and fluctuation penalty as comprehensive indicators. The total cost equals the sum of the products of electricity price, purchased power, and time step for each time period, plus the sum of the products of gas unit price, gas consumption rate, and time step. The objective function is the total cost plus the product of fluctuation penalty coefficient and fluctuation index. Constraints for maintaining the closed conservation of electrical and thermal energy are set, including power balance constraints, thermal balance constraints, and pipeline heating equations. The power balance constraint is that the power consumption of the heat pump equals the sum of purchased power and solar power output sequence. The thermal balance constraint is that the power consumption of the heat pump equals the product of the coefficient of performance, plus the product of gas consumption rate, gas calorific value, and gas boiler efficiency, plus the heat power of the thermal storage unit equals the predicted heat power of the heat load. The pipeline heating equation is that the predicted heat power of the heat load equals the product of water specific heat capacity, pipeline mass flow rate, and the difference between pipeline supply water temperature and return water temperature. A nonlinear programming approximation algorithm is used to solve the objective function and constraint equation set to obtain the time period allocation of photovoltaic-driven heat pumps, gas boilers, thermal storage heat charging and discharging, and supply water temperature.

[0011] Furthermore, the calculation of the solar power output sequence prediction reliability, the adjustment of heat pump absorption and load gap thresholds based on the prediction reliability, and the generation of multi-energy complementary heating control instructions in conjunction with time period allocation include: The prediction confidence level is generated by calculating the deviation confidence ratio between the lower quantile and the point prediction value of the future time-domain solar power output prediction sequence. The baseline heat load absorption tolerance of the temperature-controlled water tank is extracted and multiplied by the prediction confidence level to obtain the heat pump absorption threshold. The constant limit constant of the grid pre-allocated shortage is extracted and the product of the empirical adjustment factor and (1-confidence level) is subtracted to generate the load gap threshold. According to the time allocation amount in each time period and the adjusted heat pump absorption and load gap threshold, the control signal containing the continuous numerical allocation set value is output to the communication interface of each adjustable equipment cabinet and energy storage tank to generate multi-energy complementary heating control command.

[0012] In a second aspect, the present invention also provides a multi-energy complementary heating intelligent control system for a solar heating system, comprising: The calculation module is used to collect multi-source operation data of photovoltaic system and heating network to construct time series input matrix. Based on irradiance, cloud cover and heat load change rate, the time series input matrix is ​​segmented into states, and the irradiance slope, temperature difference accumulation, load gradient and state reversal index of each segment are calculated. The generation module is used to input each segment into the Long Short-Term Memory network to calculate the base value of the forget gate based on the current input and the previous hidden state. When the cloud cover changes abruptly, the base value of the forget gate is adjusted based on the continuous prediction residual characteristics to obtain the updated forget coefficient. The updated forget coefficient and the previous unit state are jointly weighted and aligned in dimension according to the dimensionless irradiance slope, temperature difference accumulation and load gradient to obtain the state retention amount. The memory compression factor is generated according to the state reversal index. The state writing amount is obtained by constraining the input gate output and candidate states. The state retention amount and the writing amount are superimposed to update the unit state and generate the solar power output sequence and fluctuation index. The adjustment module is used to construct an objective function based on the solar power output sequence, fluctuation index, heat load forecast, thermal storage status, and electricity price sequence. It solves the objective function and constraint equation set to obtain the time-period allocation of photovoltaic-driven heat pumps, gas boilers, thermal storage charging and discharging, and water supply temperature. It calculates the prediction reliability of the solar power output sequence, adjusts the heat pump absorption and load gap thresholds according to the prediction reliability, and generates multi-energy complementary heating control instructions in combination with the time-period allocation.

[0013] Preferably, the step of constructing a time-series input matrix from the multi-source operational data of the photovoltaic system and heating network includes: Historical discrete time series of photovoltaic array irradiance, module temperature, ambient temperature, cloud cover, heating network supply and return water temperature, heat load, electricity price, and state of charge of thermal storage units are acquired using data acquisition sensors. For time series with inconsistent frequencies and misalignments, the arithmetic mean interpolation of adjacent time points is used to align the data nodes and integrate the time series into a unified sampling frequency with a consistent time interval. The aligned historical data points are then stacked and merged as independent input vectors in the time domain to generate a two-dimensional time series input matrix that expresses the complete temporal development characteristics of multi-source data.

[0014] Preferably, the step of segmenting the time-series input matrix into states based on irradiance, cloud cover, and heat load change rate, and calculating the irradiance slope, cumulative temperature difference, load gradient, and state reversal index for each segment, includes: The time-series input matrix is ​​processed using a time sliding window to calculate the difference between parameters corresponding to adjacent time nodes in the time dimension, resulting in the irradiance change rate, cloud cover change rate, and heat load change rate. Waveform regions where the absolute value of the change rate continuously exceeds a set segment constant are extracted as unstable abrupt change segments. The remaining main data is truncated at a set fixed interval to form relatively static segments, thus constituting independent state segments. The irradiance slope is obtained by dividing the irradiance difference at the start and end positions of the state segment by the time span of the segment. The temperature difference accumulation is obtained by summing the absolute values ​​of the component temperature and ambient temperature differences at each moment within the state segment. The load gradient is obtained by dividing the heat power difference at the start and end positions of the state segment by the time span of the segment. The state reversal index is obtained by counting the cumulative number of sign reversals of the time-domain waveform trends of parameters within the state segment, from rising to falling or from falling to rising.

[0015] Preferably, the step of inputting each segment into the Long Short-Term Memory network to calculate the forgetting gate base value based on the current input and the previous hidden state, adjusting the forgetting gate base value based on the continuous prediction residual characteristics to obtain the updated forgetting coefficient when cloud cover changes abruptly, and jointly weighting the updated forgetting coefficient and the previous unit state according to the dimensionless irradiance slope, cumulative temperature difference, and load gradient, and aligning the dimensions to obtain the state retention amount, includes: The forget gate base value is calculated by inputting the operation matrix. The forget gate base value is equal to the sum of the product of the Sigmoid activation function applied to the correlation matrix and the concatenation result of the current input and the previous hidden state, plus the bias. When the change in cloud cover is greater than the limit threshold, the continuous prediction bias is used to intervene and obtain the updated forget coefficient. The updated forget coefficient is equal to the forget gate base value minus the product of the adjustment step difference limit rate and the difference of the normalized residual, and the value is truncated to a boundary between 0 and 1. If the change in cloud cover is not greater than the limit threshold, the updated forget coefficient is equal to the forget gate base value. The state retention is equal to the fusion vector obtained by adding the first, second, and third mapping vectors element by element. This fusion vector is then calculated by performing Hadamard products on the updated forgetting coefficient vector and the previous unit state vector. The first mapping vector is the product of the first training coefficient vector and the dimensionless irradiance slope scalar. The second mapping vector is the product of the second training coefficient vector and the temperature difference accumulation scalar. The third mapping vector is the product of the third training coefficient vector and the load gradient scalar. The first, second, and third training coefficient vectors are all learnable parameter vectors with the same dimension as the unit state.

[0016] Preferably, the step of generating a memory compression factor based on the state reversal exponent, constraining the input gate output and candidate states to obtain the state write amount, superimposing the state retention amount and the write amount to update the unit state, and generating a solar power output sequence and fluctuation index includes: A memory compression factor is generated using the natural exponential function. The memory compression factor is equal to the product of the negative decay factor constraint term of the natural exponential base and the state reversal exponent raised to the power of the power of the product. The state write amount is calculated by scaling the new variable adsorption vector using the memory compression factor. The state write amount is equal to the product of the memory compression factor and the Hadamard product of the input gate output and the candidate state. The updated unit state is then superimposed, and the unit state is equal to the sum of the state retention amount and the state write amount. Based on the updated unit state, the future time-domain solar power output sequence is generated by projecting the prediction delay execution sequence through a fully connected layer. A fluctuation index is calculated using a quadratic metric formula. The fluctuation index is equal to the square root of the sum of the squares of the differences between the power at each time step and the average power of the solar power output sequence over the total time step span.

[0017] Preferably, the step of constructing an objective function based on solar power output sequence, fluctuation index, heat load forecast, thermal storage status, and electricity price sequence, and solving the objective function and constraint equation set to obtain the time-period allocation of photovoltaic-driven heat pumps, gas boilers, thermal storage charging and discharging, and water supply temperature includes: An objective function is constructed with total cost minimization and fluctuation penalty as comprehensive indicators. The total cost equals the sum of the products of electricity price, purchased power, and time step for each time period, plus the sum of the products of gas unit price, gas consumption rate, and time step. The objective function is the total cost plus the product of fluctuation penalty coefficient and fluctuation index. Constraints for maintaining the closed conservation of electrical and thermal energy are set, including power balance constraints, thermal balance constraints, and pipeline heating equations. The power balance constraint is that the power consumption of the heat pump equals the sum of purchased power and solar power output sequence. The thermal balance constraint is that the power consumption of the heat pump equals the product of the coefficient of performance, plus the product of gas consumption rate, gas calorific value, and gas boiler efficiency, plus the heat power of the thermal storage unit equals the predicted heat power of the heat load. The pipeline heating equation is that the predicted heat power of the heat load equals the product of water specific heat capacity, pipeline mass flow rate, and the difference between pipeline supply water temperature and return water temperature. A nonlinear programming approximation algorithm is used to solve the objective function and constraint equation set to obtain the time period allocation of photovoltaic-driven heat pumps, gas boilers, thermal storage heat charging and discharging, and supply water temperature.

[0018] Preferably, the calculation of the solar power output sequence prediction reliability, adjusting the heat pump absorption and load gap thresholds based on the prediction reliability, and generating multi-energy complementary heating control instructions in conjunction with time period allocation includes: The prediction confidence level is generated by calculating the deviation confidence ratio between the lower quantile and the point prediction value of the future time-domain solar power output prediction sequence. The baseline heat load absorption tolerance of the temperature-controlled water tank is extracted and multiplied by the prediction confidence level to obtain the heat pump absorption threshold. The constant limit constant of the grid pre-allocated shortage is extracted and the product of the empirical adjustment factor and (1-confidence level) is subtracted to generate the load gap threshold. According to the time allocation amount in each time period and the adjusted heat pump absorption and load gap threshold, the control signal containing the continuous numerical allocation set value is output to the communication interface of each adjustable equipment cabinet and energy storage tank to generate multi-energy complementary heating control command.

[0019] This invention accurately detects the temporal variation patterns of multi-source operational data by segmenting the temporal input matrix into states and calculating feature indices. In a long short-term memory network, an improved forgetting gate update mechanism is used to adjust the forgetting coefficient based on continuous prediction residuals during sudden changes in cloud cover. Furthermore, multi-dimensional physical features are used to jointly generate state retention and constraint writing quantities with weighted calculations. This mechanism regulates the ratio of historical memory retention to new information integration under complex meteorological conditions, avoiding the accumulation of redundant errors and improving the prediction accuracy of solar power output. By comprehensively considering solar power output sequences, fluctuation indices, and prediction reliability to adjust equipment operating thresholds, and combining this with an objective function to solve for the time-period allocation of various heating equipment, the invention achieves synergy between photovoltaic, heat pump, gas boiler, and thermal storage systems. This improves overall energy utilization efficiency while ensuring the stable and reliable operation of the heating system. Attached Figure Description

[0020] Figure 1 A flowchart of the first embodiment; Figure 2 This is a schematic diagram of the timing curve after time alignment; Figure 3 This is a schematic diagram illustrating the photovoltaic power filtering effect. Figure 4 This is a diagram illustrating the performance comparison of the methods. Detailed Implementation

[0021] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0022] This disclosure provides an intelligent control method for multi-energy complementary heating in a solar heating system, such as... Figure 1 As shown, it includes: S1: Collect multi-source operation data of photovoltaic system and heating network to construct time series input matrix. Divide the time series input matrix into state segments according to irradiance, cloud cover and heat load change rate, and calculate the irradiance slope, temperature difference accumulation, load gradient and state reversal index of each segment.

[0023] The system acquires historical power generation, ambient temperature, and actual irradiance data of the photovoltaic array, as well as supply water temperature, return water temperature, and circulation flow rate data of the heating network through a data acquisition and monitoring control system. It uses Pandas' forward imputation function to fill in missing values ​​and Scikit-Learn's min-max scaling function to normalize the multidimensional data to the 0-1 range. A time-series input matrix is ​​constructed by stacking sliding windows along the time axis. The gradient function of NumPy is used to calculate the first derivatives of the irradiance and heat load sequences as rates of change. A time warping algorithm is employed to measure the similarity of the multidimensional feature time series. Finally, the KMeans clustering algorithm is used to segment the time-series input matrix into independent segments, such as stationary and fluctuating periods, using cloud cover abrupt change points and extreme points of rate of change as boundaries. For each segment, the linear regression function of the scientific computing library Scipy is used to fit the irradiance time series to obtain the slope of the straight line as the irradiance slope. The trapezoidal integral function is used to calculate the definite integral of the difference series between the component temperature and the ambient temperature along the time step to obtain the cumulative temperature difference. The difference function is used to calculate the difference in heat load between adjacent time steps as the load gradient. The ratio of the number of sign flips of the first derivative of irradiance within the segment to the segment length is calculated as the state reversal index.

[0024] In an alternative embodiment, the step of constructing a time-series input matrix from the multi-source operational data of the photovoltaic system and heating network includes: Historical discrete time series of photovoltaic array irradiance, module temperature, ambient temperature, cloud cover, heating network supply and return water temperature, heat load, electricity price, and state of charge of thermal storage units are acquired using data acquisition sensors. For time series with inconsistent frequencies and misalignments, the arithmetic mean interpolation of adjacent time points is used to align the data nodes and integrate the time series into a unified sampling frequency with a consistent time interval. The aligned historical data points are then stacked and merged as independent input vectors in the time domain to generate a two-dimensional time series input matrix that expresses the complete temporal development characteristics of multi-source data.

[0025] The acquisition of multi-source operational data relies on a high-precision sensor network deployed in the physical system. For example, a total radiation meter is used to collect photovoltaic array irradiance measurements ranging from 0 to 2000 W / m², and a PT100 platinum resistance temperature sensor is used to collect module and ambient temperature measurements ranging from -40°C to 85°C. The SCADA system is used to obtain heat load in kilowatts and thermal storage SOC ranging from 0 to 100%. Due to differences in data reporting frequencies between different subsystems (e.g., photovoltaic inverter data is typically on a 5-minute timescale, while heating network SCADA data may be on a 15-minute timescale or electricity price data on an hourly timescale), a globally unified baseline sampling step size, preferably 15 minutes, is set when performing time series alignment. For timestamps with data misalignment or missing data due to communication delays, an arithmetic mean interpolation method is used. For example, if module temperature data at 10:15 is missing, the arithmetic mean of the values ​​collected at 10:10 and 10:20 is extracted for smoothing. After forced alignment of data nodes, the above parameters are extracted for each time step to construct a one-dimensional row vector containing 9-dimensional feature parameters. By tracing back and extracting historical sequence data from the past 24 hours, which includes 96 15-minute time steps, these 96 independent one-dimensional feature vectors are sequentially stacked downwards in the time domain to construct a two-dimensional temporal input tensor matrix with a data dimension of 96×9. This matrix serves as the standard temporal base for feature extraction in the Long Short-Term Memory network. The time-aligned time-series curves of multi-source operational data such as photovoltaic irradiance, ambient temperature, and heat load are shown below. Figure 2 As shown.

[0026] In one possible embodiment, the step of segmenting the time-series input matrix into states based on irradiance, cloud cover, and heat load change rate, and calculating the irradiance slope, cumulative temperature difference, load gradient, and state reversal index for each segment, includes: The time-series input matrix is ​​processed using a time sliding window to calculate the difference between parameters corresponding to adjacent time nodes in the time dimension, resulting in the irradiance change rate, cloud cover change rate, and heat load change rate. Waveform regions where the absolute value of the change rate continuously exceeds a set segment constant are extracted as unstable abrupt change segments. The remaining main data is truncated at a set fixed interval to form relatively static segments, thus constituting independent state segments. The irradiance slope is obtained by dividing the irradiance difference at the start and end positions of the state segment by the time span of the segment. The temperature difference accumulation is obtained by summing the absolute values ​​of the component temperature and ambient temperature differences at each moment within the state segment. The load gradient is obtained by dividing the heat power difference at the start and end positions of the state segment by the time span of the segment. The state reversal index is obtained by counting the cumulative number of sign reversals of the time-domain waveform trends of parameters within the state segment, from rising to falling or from falling to rising.

[0027] When segmenting the state of the two-dimensional time-series input matrix, a discrete-time sliding window of length 2 is set, and the global data is traversed with a sliding step size of 1 time step. Preset threshold values ​​for high-frequency segmentation constants for each parameter: the irradiance change rate constant is set to 50 W / m² every 15 minutes, the cloud cover change rate constant is set to 20% every 15 minutes, and the heat load change rate constant is set to 15 kW every 15 minutes. When the absolute value of the first-order difference rate of change of any of the above parameters exceeds the set threshold for three consecutive time steps (45 minutes), the time domain boundary of this drastic fluctuation is automatically truncated and divided into unstable abrupt change segments. For normal stable period data with smooth waveforms and change rates below the threshold, uniform truncation is performed with a fixed span of 2 hours (8 consecutive time steps), thus forming an independent set of state segments together with the abrupt change segments.

[0028] Geometric and physical features are extracted from each segment of the divided state. If a segment spans 60 minutes and includes four discrete step lengths, with irradiances of 300 W / m² and 700 W / m² at the start and end points respectively, the irradiance slope is calculated to be 6.67 W / (m²·min). The measured temperatures of the photovoltaic modules and the ambient temperature at the four time points within the segment are extracted, and the absolute values ​​of the temperature differences at each step are calculated, such as 5℃, 7℃, 6℃, and 8℃ respectively. The cumulative temperature difference is obtained by summing these absolute values ​​as 26℃. The load gradient is obtained by subtracting the initial demand from the heat load demand at the end of the segment from the initial demand, such as 80 kW divided by the 60-minute span, resulting in an average load ramp rate of 1.33 kW / min. A sign discrimination function is used to detect the sign of the first derivative within the segment. If the irradiance waveform experiences an increase followed by a decrease followed by an increase, and the first derivative sign alternates twice, the corresponding state reversal index is assigned an integer value of 2.

[0029] S2. Input each segment into the Long Short-Term Memory network and calculate the forget gate base value based on the current input and the previous hidden state. When cloud cover changes abruptly, adjust the forget gate base value based on the continuous prediction residual characteristics to obtain the updated forget coefficient. Based on the dimensionless irradiance slope, temperature difference accumulation and load gradient, jointly weight the updated forget coefficient and the previous unit state and align the dimensions to obtain the state retention amount. Generate the memory compression factor based on the state reversal index. Constrain the input gate output and candidate states to obtain the state writing amount. Superimpose the state retention amount and the writing amount to update the unit state and generate the solar power output sequence and fluctuation index.

[0030] A Long Short-Term Memory (LSTM) network model based on the PyTorch deep learning framework is constructed. Each segmented sequence is input step-by-step. A linear mapping layer and a Sigmoid nonlinear activation function are used to perform matrix multiplication and bias addition operations on the current input vector and the hidden state vector from the previous time step, outputting a forgetting gate base value in the range of 0 to 1. When the difference in cloud cover between adjacent time steps exceeds a set threshold, indicating a sudden change in cloud cover, the absolute error sequence between the predicted and actual values ​​from three historical time steps is extracted. An error penalty coefficient is generated by mapping this error using a multilayer perceptron. The penalty coefficient is then multiplied by the forgetting gate base value using a Hadamard product to obtain the updated forgetting coefficient.

[0031] The data scaling function is called to perform zero-mean normalization mapping on the cumulative irradiance slope temperature difference and load gradient. These three are concatenated into a physical feature vector, which is then input into a fully connected layer and mapped to the same dimension as the unit state. The dimension-aligned physical feature vector is updated with the forgetting coefficient, and the previous unit state is multiplied element-wise to obtain the state retention. The natural exponential function is used to calculate a memory compression factor between 0 and 1, using the negative of the state reversal exponent as the exponent. This memory compression factor is multiplied by the product matrix of the input gate activation value and the candidate state to filter information and obtain the state writing amount. The state retention amount and the state writing amount are added element-wise to update the current unit state. The current hidden state is generated through the output gate and the Tanh hyperbolic tangent activation function. The solar power output prediction sequence is decoded by a fully connected regression layer, and the standard deviation of the power output prediction sequence is calculated using the sliding window standard deviation function as a fluctuation index. The original photovoltaic power output fluctuation curve is compared with the power output curve predicted and filtered by the model of this invention. Figure 3 As shown.

[0032] In one possible embodiment, the step of inputting each segment into a long short-term memory network to calculate a base value for the forgetting gate based on the current input and the previous hidden state, adjusting the base value for the forgetting gate based on the continuous prediction residual characteristics to obtain an updated forgetting coefficient when cloud cover changes abruptly, and jointly weighting the updated forgetting coefficient and the previous unit state according to the dimensionless irradiance slope, cumulative temperature difference, and load gradient, and aligning the dimensions to obtain the state retention amount includes: The forget gate base value is calculated by inputting the operation matrix. The forget gate base value is equal to the sum of the product of the Sigmoid activation function applied to the correlation matrix and the concatenation result of the current input and the previous hidden state, plus the bias. When the change in cloud cover is greater than the limit threshold, the continuous prediction bias is used to intervene and obtain the updated forget coefficient. The updated forget coefficient is equal to the forget gate base value minus the product of the adjustment step difference limit rate and the difference of the normalized residual, and the value is truncated to a boundary between 0 and 1. If the change in cloud cover is not greater than the limit threshold, the updated forget coefficient is equal to the forget gate base value. The state retention is equal to the fusion vector obtained by adding the first, second, and third mapping vectors element by element. This fusion vector is then calculated by performing Hadamard products on the updated forgetting coefficient vector and the previous unit state vector. The first mapping vector is the product of the first training coefficient vector and the dimensionless irradiance slope scalar. The second mapping vector is the product of the second training coefficient vector and the temperature difference accumulation scalar. The third mapping vector is the product of the third training coefficient vector and the load gradient scalar. The first, second, and third training coefficient vectors are all learnable parameter vectors with the same dimension as the unit state.

[0033] The Long Short-Term Memory (LSTM) network structure includes an input layer, hidden layers, and a fully connected output layer. The input layer receives external time-series data, the hidden layers extract features through internal recurrent cell states and gating mechanisms, and the fully connected output layer maps the hidden layer states to prediction results. The input to the LTM network is a concatenated two-dimensional time-series input matrix, and the output is the mapped solar power output sequence. The formula for calculating the base value of the forgetting gate within the hidden layers is f = ... ×(W×x+b), where f is the base value of the forget gate. Let W be the activation function, W be the weight association matrix, x be the concatenated vector of the previous hidden state and the current input features, and b be the intercept bias. During the internal flow control computation phase of the Long Short-Term Memory (LSTM) network architecture, the optimal dimension of the unit nodes of the hidden layer neurons is 64. The forget gate base value vector is calculated by concatenating the extracted current input feature tensor vector with the 64-dimensional hidden state vector returned from the previous time step along the axis. The concatenated vector is then multiplied by the association matrix initially assigned weights by the network. After adding the intercept bias, it is nested into the Sigmoid function to smoothly compress and normalize all forget gate base values ​​to a 0-1 data space.

[0034] To address the uncontrolled forecasting caused by sudden cloudy weather, a threshold of 30% is set for the instantaneous change in cloud cover. When a sudden surge of overcast area exceeding 30% is detected between two steps, a feedforward bias adjustment intervention command is immediately triggered. Under this intervention mechanism, the control flow extracts the absolute residual of the photovoltaic output point prediction from the previous moment (e.g., an error of 15kW) and the residual from the moment before that (i.e., two moments prior) (e.g., 5kW), calculating a difference of 10kW between the two residuals. This difference is divided by a preset baseline power, e.g., 20kW, yielding a normalized residual difference of 0.5. A smoothing adjustment step difference limit rate hyperparameter is preferably selected as 0.1, and a penalty term of 0.05 is subtracted from the baseline value. An intermediate value function is used to hard-truncate the deduction result at upper and lower limits to prevent overflow, thus obtaining the updated forgetting coefficient after adjustment feedback. If the 30% threshold is not exceeded, the original baseline value is used.

[0035] Max-min normalization is used to uniformly compress and map three heterogeneous scalar quantities—irradiance slope, cumulative temperature difference, and load gradient—to the 0-1 domain. Simultaneously, three 64×1 dimensional learnable training coefficient column vector matrices are pre-constructed. The first training coefficient vector is multiplied by the normalized irradiance slope scalar to obtain the first mapping vector; the second training coefficient vector is multiplied by the cumulative temperature difference scalar to obtain the second mapping vector; and the third training coefficient vector is multiplied by the load gradient scalar to obtain the third mapping vector. These three 64-dimensional mapping vectors are then element-wise summed to obtain a fusion vector. This fusion vector, along with the previously calculated update forgetting coefficient vector, is then multiplied element-wise by the cell state vector of the previous unit of the network—this is called Hadamard product calculation. After filtering out redundant historical interference, the state retention vector with the matching dimension is extracted.

[0036] In one possible embodiment, the step of generating a memory compression factor based on the state inversion exponent, constraining the input gate output and candidate states to obtain the state write amount, superimposing the state retention amount and the write amount to update the unit state, and generating a solar power output sequence and fluctuation index includes: A memory compression factor is generated using the natural exponential function. The memory compression factor is equal to the product of the negative decay factor constraint term of the natural exponential base and the state reversal exponent raised to the power of the power of the product. The state write amount is calculated by scaling the new variable adsorption vector using the memory compression factor. The state write amount is equal to the product of the memory compression factor and the Hadamard product of the input gate output and the candidate state. The updated unit state is then superimposed, and the unit state is equal to the sum of the state retention amount and the state write amount. Based on the updated unit state, the future time-domain solar power output sequence is generated by projecting the prediction delay execution sequence through a fully connected layer. A fluctuation index is calculated using a quadratic metric formula. The fluctuation index is equal to the square root of the sum of the squares of the differences between the power at each time step and the average power of the solar power output sequence over the total time step span.

[0037] The optimal tuning range for the negative decay factor constraint in the preset network parameters is between 0.1 and 0.5, typically a constant value of 0.2. If the state reversal index within the current segment is 4, indicating four oscillations, then calculate e. -0.8 The memory compression factor scalar is found to be approximately 0.449 at this point. A gating combination operation is initiated, where the newly calculated network input gate output feature vector is multiplied element-wise with the internal candidate state vector generated through hyperbolic tangent activation to obtain a new variable adsorption vector. This new variable adsorption vector is then multiplied by the scalar 0.449 for linear compression, resulting in a purified and denoised state write vector.

[0038] Then, this state write quantity is combined with the state retention quantity obtained in the previous steps through corresponding node addition and synthesis operations to obtain the latest update unit state cell vector that takes into account both stable memory and resistance to drastic disturbances. This latest vector is then connected to the prediction decoding layer, where the number of nodes is a linear fully connected layer network with a prediction horizon of time steps. For example, if the projection window is set to the next 4 hours, covering 16 forward time steps, the fully connected layer maps the projection window to a continuous projected output sequence with 16 discrete node elements, representing the output of the next 16 consecutive solar power prediction values ​​in kW. The module calls the discrete statistical quadratic metric variance logic in parallel to calculate the mean power of the 16 output data. The output power of each step is subtracted from this mean value, squared, and the absolute values ​​of the 16 quadratic variances are algebraically accumulated. The sum is divided by the total step length constant 16, and the arithmetic square root is taken. The resulting root mean square deviation of the power standard deviation is used as a core fluctuation indicator to judge and transmit the strength of discrete fluctuations in the photovoltaic prediction array output to the subsequent allocation layer.

[0039] S3 constructs an objective function based on solar power output sequence, fluctuation index, heat load forecast, thermal storage status, and electricity price sequence. It solves the objective function and constraint equation set to obtain the time-period allocation of photovoltaic-driven heat pumps, gas boilers, thermal storage charging and discharging, and water supply temperature. It calculates the prediction reliability of the solar power output sequence, adjusts the heat pump absorption and load gap thresholds according to the prediction reliability, and generates multi-energy complementary heating control instructions in combination with the time-period allocation.

[0040] A nonlinear programming model is constructed with the goal of minimizing overall operating costs. The objective function includes the electricity consumption cost of the photovoltaic-driven heat pump, the fuel cost of the gas-fired boiler, the operation and maintenance depreciation cost of the thermal storage equipment, and the curtailment penalty. Inputs include the solar power output sequence fluctuation index, the predicted heat load, the current state of charge of the thermal storage tank, and the time-of-use electricity price sequence. A set of constraint equations is established, including supply and demand power heat balance constraints, equipment ramp-up rate constraints, upper and lower limits of thermal storage capacity constraints, and the comfort range of the supply water temperature. The Gurobi branch and bound algorithm is used in conjunction with the nonlinear programming solver to iteratively find the optimal solution. The solution result matrix is ​​analyzed to obtain the power input of the heat pump, the heat generation of the gas-fired boiler, the charging and discharging heat flow rate of the thermal storage tank, and the setpoint of the supply water temperature of the heat exchange station for each scheduling period, which are used as the time period allocation. A normalized negative exponential function is used to inversely map the fluctuation index of the acquired solar power output sequence to generate a prediction confidence level between zero and one. A baseline threshold parameter is set, and the prediction confidence level is multiplied by the baseline threshold to obtain a decay coefficient. This coefficient is used to lower the original maximum absorption capacity of the heat pump, and the load gap trigger threshold is adjusted inversely according to the confidence level. That is, the lower the prediction confidence level, the lower the threshold, making it easier for the backup heat source to start and fill potential gaps in advance. The adjusted operating threshold is integrated with the above time period allocation, and the PyModbus communication toolkit is used to encode data according to the ModbusTCP industrial control protocol format to generate multi-energy complementary heating control commands. These commands are then sent to the programmable logic controller via industrial Ethernet to execute the physical equipment's action adjustment.

[0041] In one possible embodiment, the step of constructing an objective function based on solar power output sequence, fluctuation index, heat load forecast, thermal storage status, and electricity price sequence, and solving the objective function and constraint equation set to obtain the time-period allocation of photovoltaic-driven heat pumps, gas boilers, thermal storage charging and discharging, and water supply temperature includes: An objective function is constructed with total cost minimization and fluctuation penalty as comprehensive indicators. The total cost equals the sum of the products of electricity price, purchased power, and time step for each time period, plus the sum of the products of gas unit price, gas consumption rate, and time step. The objective function is the total cost plus the product of fluctuation penalty coefficient and fluctuation index. Constraints for maintaining the closed conservation of electrical and thermal energy are set, including power balance constraints, thermal balance constraints, and pipeline heating equations. The power balance constraint is that the power consumption of the heat pump equals the sum of purchased power and solar power output sequence. The thermal balance constraint is that the power consumption of the heat pump equals the product of the coefficient of performance, plus the product of gas consumption rate, gas calorific value, and gas boiler efficiency, plus the heat power of the thermal storage unit equals the predicted heat power of the heat load. The pipeline heating equation is that the predicted heat power of the heat load equals the product of water specific heat capacity, pipeline mass flow rate, and the difference between pipeline supply water temperature and return water temperature. A nonlinear programming approximation algorithm is used to solve the objective function and constraint equation set to obtain the time period allocation of photovoltaic-driven heat pumps, gas boilers, thermal storage heat charging and discharging, and supply water temperature.

[0042] For each discrete scheduling period in the future, by time step Using 0.25 hours as a baseline, the time-of-use electricity price (e.g., from a trough of 0.3 yuan / kWh to a peak of 1.2 yuan / kWh) is extracted and multiplied by the real-time power purchase capacity and time step of the main grid to calculate the electricity purchase cost. Simultaneously, the fuel cost is calculated by multiplying the fixed natural gas price (typically 3 yuan / m³) by the gas consumption rate and time step. The two costs are added together to form the explicit total economic cost. A preset system fluctuation penalty coefficient weight is retrieved, preferably set to a standby capacity conversion rate of 2.5 yuan / kW. This coefficient is multiplied by the scalar of the photovoltaic power fluctuation index obtained in the previous step to convert it into an equivalent penalty compensation term. The total economic cost plus the equivalent penalty term is combined to form the evaluation objective function used for algorithm optimization to minimize the extreme value. When constructing the scheduling constraint model set, a bilateral equal coupling constraint for electric and thermal power balance is established, requiring the total power consumption of the heat pump equipment to be equal to the sum of the power supplied by the external main grid and the predicted solar power output generated by the self-generated photovoltaic array. For the heat conduction balance condition, the rated parameters of the equipment are used to calibrate the heating capacity. The normal average value of the heat pump performance coefficient that varies with temperature is extracted and set to 3.5. Multiplying the power consumption to obtain the heat source on the foundation side, the fixed constant of the low heating value of natural gas (9.7 kW·h / m³) and the rated conversion efficiency of the gas boiler unit (set to 0.92) are read and multiplied by the consumption rate to obtain the supplementary heat source on the second side. The two sets of heat generation are added together and the bidirectional charging and discharging heat compensation adjustment power of the water tank heat storage unit is added. The release is recorded as positive and the storage is recorded as negative. This ensures that the system's synthesized output is equal to the sum of the predicted heat power of the load given at the end of the system. Combining the water pipeline transportation equation, the aforementioned heat load heat power, the specific heat capacity constant of the medium water in the pipe (4.18 kJ / (kg·℃)), the real-time pipeline mass flow rate driven by the circulating pump (e.g., rated 10 kg / s), and the physical difference between the secondary network supply water temperature and the return water set temperature required by the control stage are further multiplied into a closed-loop equation. Because the optimization system includes binary discrete logic for the start-up and shutdown of gas-fired boilers and other units, as well as continuous mathematical boundaries for fluid allocation, and the constraint equations contain nonlinear terms involving variable multiplication, a solver module supporting nonlinear programming is invoked. An algorithm combining branch-and-bound mechanism and generalized simplified gradient method is used to globally optimize the model equation set. Within the constraints, the optimal joint time-period allocation decision matrix table covering the photovoltaic heat pump power consumption, gas regulating valve set gas supply rate, water tank energy storage and release curve, and the target supply water temperature of the primary and secondary network plate heat exchangers is successfully calculated.

[0043] In one possible embodiment, the calculation of the predicted reliability of the solar power output sequence, the adjustment of the heat pump absorption and load gap thresholds based on the predicted reliability, and the generation of multi-energy complementary heating control instructions in conjunction with the time-period allocation include: The prediction confidence level is generated by calculating the deviation confidence ratio between the lower quantile and the point prediction value of the future time-domain solar power output prediction sequence. The baseline heat load absorption tolerance of the temperature-controlled water tank is extracted and multiplied by the prediction confidence level to obtain the heat pump absorption threshold. The constant limit constant of the grid pre-allocated shortage is extracted and the product of the empirical adjustment factor and (1-confidence level) is subtracted to generate the load gap threshold. According to the time allocation amount in each time period and the adjusted heat pump absorption and load gap threshold, the control signal containing the continuous numerical allocation set value is output to the communication interface of each adjustable equipment cabinet and energy storage tank to generate multi-energy complementary heating control command.

[0044] During the risk assessment phase of source-side prediction uncertainty, the control unit concurrently calls the quantile regression model algorithm to extract the 5% extreme lower quantile of the photovoltaic power output distribution for the corresponding future time period. When the network calculates the predicted output of the future benchmark photovoltaic point to be 100kW, if the model measures the lower quantile's pessimistic limit to be 90kW, the relative absolute deviation between the two is calculated, i.e., a deviation of 10kW. This difference is divided by the predicted point value to obtain a deviation rate of 10%, or 0.1. Subtracting this deviation ratio from 1 yields a dimensionless constant of 0.9, which is used to define the prediction confidence level representing the confidence level of this time-series node. Based on this scalar, to prevent sudden photovoltaic drops from causing the unit to disconnect from the grid beyond its limits, the main control unit extracts the basic heat load absorption tolerance base established in the factory safety guidelines for the phase change temperature-regulating hot water storage tank, and extracts the nominal buffer pool buffer thermal energy parameter of the hot water tank, such as 50kW. Multiplying 50kW by the aforementioned confidence level of 0.9, the upper limit threshold for heat pump absorption safety is reduced to 45kW. The SCADA platform within the plant retrieves the grid pre-configured shortage constant safety limit constant set to cope with sudden power outages, which is generally rigidly assigned a value of 20kW. The threshold is lowered by subtracting an empirical adjustment factor, such as 10, from the limit constant and multiplying it by the confidence complement 1-0.9. This derives a flexible and lowered anti-disturbance load gap early warning alarm threshold of 19kW at this moment, thereby realizing an active defense logic where the threshold decreases synchronously when the prediction confidence decreases and the backup heat source is more easily triggered.

[0045] During the scheduling and distribution process at the edge layer, the on-site integrated energy controller merges the time-period allocation table sequence data, which includes water supply temperature and power frequency bands and was previously obtained through nonlinear optimization, with the newly revised and determined upper limit for heat pump absorption and lower limit for power shortage thresholds. Through data encapsulation and packaging at the industrial control computer hardware layer, and along the industrial LAN backbone at the factory's bottom layer—such as an Ethernet network based on the Modbus protocol and an RS485 twisted-pair communication architecture for the underlying actuators—a series of continuous analog and digital control signals covering 0 to 100% equipment power control, 0 to 50Hz operating frequency adjustment, and 4 to 20mA analog opening signals are sent to the integrated photovoltaic-inverter cabinet, the variable frequency air source heat pump PLC, and the energy storage tank valve actuator with a stepper servo motor. This achieves millisecond-level, multi-energy complementary, coordinated heating joint control at the bottom layer.

[0046] The experimental data consisted of one year's worth of continuous operation monitoring data from a real multi-energy complementary microgrid demonstration project, covering multiple time-series parameters such as photovoltaic array irradiance, component temperature, and heating network status. The data sampling frequency step size was uniformly set to 15 minutes, and the time-series input matrix was divided into training, validation, and test sets according to time sequence. A high-performance tensor core accelerator computing card and a deep learning computing framework were used to construct the validation model. The evaluation benchmarks were the root mean square error of photovoltaic output sequence prediction and the average daily comprehensive operation and scheduling cost of the multi-energy complementary microgrid. Four comparison schemes were established, including a basic multilayer perceptron network group, a conventional long short-term memory network group, an ablation group that removed cloud cover, residual forgetting, update coefficient adjustment, and state writing inversion compression constraints, and the complete algorithm scheme of this invention, which includes all underlying flow control adjustment features and a nonlinear hybrid optimization mechanism.

[0047] After equal iterations and independent testing of all schemes, the root mean square error (RMSE) of photovoltaic (PV) power prediction measured by the basic multilayer perceptron network reached as high as 18.55 kW, with a corresponding daily total operating cost of 1456.2 yuan for the heating microgrid system. The conventional long short-term memory (LSTM) network model reduced the RMSE of PV power prediction to 14.12 kW, lowering the daily total operating cost of the microgrid to 1243.8 yuan. The ablation group model, leveraging its coefficient weight alignment capability, reduced the RMSE of prediction to 9.84 kW, reducing the daily microgrid system cost to 1042.5 yuan. The complete scheme model of this invention exhibits optimal performance. Measurements show that the root mean square error (RMSE) of the photovoltaic output sequence prediction under complex environmental disturbances is significantly reduced to only 5.26 kW. The optimal daily average scheduling cost of microgrid complementarity, output under the joint scheduling mechanism of the complex constraint model set optimization, is compressed to an extremely low 875.4 yuan. The bar chart comparing the RMSE of photovoltaic output prediction and the daily average operating cost of the microgrid with MLP, conventional LSTM, ablation model, and the complete scheme of this invention is shown below. Figure 4 As shown, the results indicate that the proposed solution is superior to other methods in both prediction accuracy and system economy.

[0048] This study verifies that under severe boundary conditions of rapidly changing cloud cover, relying solely on feature weighting cannot prevent the backward propagation of historical distortion residual errors between time series. This invention utilizes a memory compression feature operator and a prediction bias hard truncation compensation strategy to adjust the write retention share of the hidden layer in the cell unit, successfully eliminating high-frequency waveform oscillation noise at micro-nodes. The resulting pre-smooth output base-assisted heating optimization equation avoids the penalty overhead caused by blindly starting gas boilers or ineffective phase change water tank throughput, thus achieving the optimal solution for lean energy scheduling in the microgrid.

[0049] This disclosure also provides a multi-energy complementary heating intelligent control system for a solar heating system, including: The calculation module is used to collect multi-source operation data of photovoltaic system and heating network to construct time series input matrix. Based on irradiance, cloud cover and heat load change rate, the time series input matrix is ​​segmented into states, and the irradiance slope, temperature difference accumulation, load gradient and state reversal index of each segment are calculated. The generation module is used to input each segment into the Long Short-Term Memory network to calculate the base value of the forget gate based on the current input and the previous hidden state. When the cloud cover changes abruptly, the base value of the forget gate is adjusted based on the continuous prediction residual characteristics to obtain the updated forget coefficient. The updated forget coefficient and the previous unit state are jointly weighted and aligned in dimension according to the dimensionless irradiance slope, temperature difference accumulation and load gradient to obtain the state retention amount. The memory compression factor is generated according to the state reversal index. The state writing amount is obtained by constraining the input gate output and candidate states. The state retention amount and the writing amount are superimposed to update the unit state and generate the solar power output sequence and fluctuation index. The adjustment module is used to construct an objective function based on the solar power output sequence, fluctuation index, heat load forecast, thermal storage status, and electricity price sequence. It solves the objective function and constraint equation set to obtain the time-period allocation of photovoltaic-driven heat pumps, gas boilers, thermal storage charging and discharging, and water supply temperature. It calculates the prediction reliability of the solar power output sequence, adjusts the heat pump absorption and load gap thresholds according to the prediction reliability, and generates multi-energy complementary heating control instructions in combination with the time-period allocation.

[0050] In this specification, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise limited, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element. In this document, "a," "an," "the," "the," and "its" may also include plural forms unless the context clearly indicates otherwise. "Multiple" refers to at least two, such as 2, 3, 5, or 8, etc. "And / or" includes any and all combinations of the associated listed items.

[0051] The various embodiments in this specification are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. The various embodiments can be combined as needed, and the same or similar parts can be referred to each other.

[0052] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A multi-energy complementary heating intelligent control method for a solar heating system, characterized in that, include: A time-series input matrix is ​​constructed by collecting multi-source operation data of photovoltaic system and heating network. The time-series input matrix is ​​segmented into states according to irradiance, cloud cover and heat load change rate, and the irradiance slope, temperature difference accumulation, load gradient and state reversal index of each segment are calculated. Each segment is input into the Long Short-Term Memory Network to calculate the base value of the forget gate based on the current input and the previous hidden state. When cloud cover changes abruptly, the base value of the forget gate is adjusted based on the continuous prediction residual characteristics to obtain the updated forgetting coefficient. The updated forgetting coefficient and the previous unit state are jointly weighted and aligned in dimension according to the dimensionless irradiance slope, temperature difference accumulation and load gradient to obtain the state retention amount. The memory compression factor is generated according to the state reversal index. The state writing amount is obtained by constraining the input gate output and candidate states. The state retention amount and the writing amount are superimposed to update the unit state and generate the solar power output sequence and fluctuation index. An objective function is constructed based on the solar power output sequence, fluctuation index, heat load forecast, thermal storage status, and electricity price sequence. The objective function and constraint equation set are solved to obtain the time-period allocation of photovoltaic-driven heat pumps, gas boilers, thermal storage charging and discharging, and water supply temperature. The prediction reliability of the solar power output sequence is calculated. The heat pump absorption and load gap thresholds are adjusted according to the prediction reliability. Multi-energy complementary heating control instructions are generated by combining the time-period allocation.

2. The method according to claim 1, characterized in that, The process of collecting multi-source operational data from photovoltaic systems and heating networks to construct a time-series input matrix includes: Historical discrete time series of photovoltaic array irradiance, module temperature, ambient temperature, cloud cover, heating network supply and return water temperature, heat load, electricity price, and state of charge of thermal storage units are acquired using data acquisition sensors. For time series with inconsistent frequencies and misalignments, the arithmetic mean interpolation of adjacent time points is used to align the data nodes and integrate the time series into a unified sampling frequency with a consistent time interval. The aligned historical data points are then stacked and merged as independent input vectors in the time domain to generate a two-dimensional time series input matrix that expresses the complete temporal development characteristics of multi-source data.

3. The method according to claim 2, characterized in that, The process of segmenting the time-series input matrix based on irradiance, cloud cover, and heat load change rate, and calculating the irradiance slope, cumulative temperature difference, load gradient, and state reversal index for each segment, includes: The time-series input matrix is ​​processed using a time sliding window to calculate the difference between parameters corresponding to adjacent time nodes in the time dimension, resulting in the irradiance change rate, cloud cover change rate, and heat load change rate. Waveform regions where the absolute value of the change rate continuously exceeds a set segment constant are extracted as unstable abrupt change segments. The remaining main data is truncated at a set fixed interval to form relatively static segments, thus constituting independent state segments. The irradiance slope is obtained by dividing the irradiance difference at the start and end positions of the state segment by the time span of the segment. The temperature difference accumulation is obtained by summing the absolute values ​​of the component temperature and ambient temperature differences at each moment within the state segment. The load gradient is obtained by dividing the heat power difference at the start and end positions of the state segment by the time span of the segment. The state reversal index is obtained by counting the cumulative number of sign reversals of the time-domain waveform trends of parameters within the state segment, from rising to falling or from falling to rising.

4. The method according to claim 3, characterized in that, The process involves inputting each segment into a Long Short-Term Memory (LSTM) network to calculate a base value for the forgetting gate based on the current input and the previous hidden state. When cloud cover changes abruptly, the base value for the forgetting gate is adjusted based on the characteristics of the continuous prediction residuals to obtain an updated forgetting coefficient. The updated forgetting coefficient and the previous unit state are jointly weighted and aligned in dimension based on the dimensionless irradiance slope, cumulative temperature difference, and load gradient to obtain the state retention amount, including: The forget gate base value is calculated by inputting the operation matrix. The forget gate base value is equal to the sum of the product of the Sigmoid activation function applied to the correlation matrix and the concatenation result of the current input and the previous hidden state, plus the bias. When the change in cloud cover is greater than the limit threshold, the continuous prediction bias is used to intervene and obtain the updated forget coefficient. The updated forget coefficient is equal to the forget gate base value minus the product of the adjustment step difference limit rate and the difference of the normalized residual, and the value is truncated to a boundary between 0 and 1. If the change in cloud cover is not greater than the limit threshold, the updated forget coefficient is equal to the forget gate base value. The state retention is equal to the fusion vector obtained by adding the first, second, and third mapping vectors element by element. This fusion vector is then calculated by performing Hadamard products on the updated forgetting coefficient vector and the previous unit state vector. The first mapping vector is the product of the first training coefficient vector and the dimensionless irradiance slope scalar. The second mapping vector is the product of the second training coefficient vector and the temperature difference accumulation scalar. The third mapping vector is the product of the third training coefficient vector and the load gradient scalar. The first, second, and third training coefficient vectors are all learnable parameter vectors with the same dimension as the unit state.

5. The method according to claim 1, characterized in that, The process of generating a memory compression factor based on the state reversal index, constraining the input gate output and candidate states to obtain the state write amount, and superimposing the state retention amount and the write amount to update the unit state, generates a solar power output sequence and fluctuation index, including: A memory compression factor is generated using the natural exponential function. The memory compression factor is equal to the product of the negative decay factor constraint term of the natural exponential base and the state reversal exponent raised to the power of the power of the product. The state write amount is calculated by scaling the new variable adsorption vector using the memory compression factor. The state write amount is equal to the product of the memory compression factor and the Hadamard product of the input gate output and the candidate state. The updated unit state is then superimposed, and the unit state is equal to the sum of the state retention amount and the state write amount. Based on the updated unit state, the future time-domain solar power output sequence is generated by projecting the prediction delay execution sequence through a fully connected layer. A fluctuation index is calculated using a quadratic metric formula. The fluctuation index is equal to the square root of the sum of the squares of the differences between the power at each time step and the average power of the solar power output sequence over the total time step span.

6. The method according to claim 1, characterized in that, The objective function is constructed based on solar power output sequence, fluctuation index, heat load forecast, thermal storage status, and electricity price sequence. Solving the objective function and constraint equation set yields the time-period allocation of photovoltaic-driven heat pumps, gas boilers, thermal storage charging and discharging, and water supply temperature, including: An objective function is constructed with total cost minimization and fluctuation penalty as comprehensive indicators. The total cost equals the sum of the products of electricity price, purchased power, and time step for each time period, plus the sum of the products of gas unit price, gas consumption rate, and time step. The objective function is the total cost plus the product of fluctuation penalty coefficient and fluctuation index. Constraints for maintaining the closed conservation of electrical and thermal energy are set, including power balance constraints, thermal balance constraints, and pipeline heating equations. The power balance constraint is that the power consumption of the heat pump equals the sum of purchased power and solar power output sequence. The thermal balance constraint is that the power consumption of the heat pump equals the product of the coefficient of performance, plus the product of gas consumption rate, gas calorific value, and gas boiler efficiency, plus the heat power of the thermal storage unit equals the predicted heat power of the heat load. The pipeline heating equation is that the predicted heat power of the heat load equals the product of water specific heat capacity, pipeline mass flow rate, and the difference between pipeline supply water temperature and return water temperature. A nonlinear programming approximation algorithm is used to solve the objective function and constraint equation set to obtain the time period allocation of photovoltaic-driven heat pumps, gas boilers, thermal storage heat charging and discharging, and supply water temperature.

7. The method according to claim 1, characterized in that, The calculation of the solar power output sequence prediction reliability, the adjustment of heat pump absorption and load gap thresholds based on the prediction reliability, and the generation of multi-energy complementary heating control instructions in conjunction with time period allocation include: The prediction confidence level is generated by calculating the deviation confidence ratio between the lower quantile and the point prediction value of the future time-domain solar power output prediction sequence. The baseline heat load absorption tolerance of the temperature-controlled water tank is extracted and multiplied by the prediction confidence level to obtain the heat pump absorption threshold. The constant limit constant of the grid pre-allocated shortage is extracted and the product of the empirical adjustment factor and (1-confidence level) is subtracted to generate the load gap threshold. According to the time allocation amount in each time period and the adjusted heat pump absorption and load gap threshold, the control signal containing the continuous numerical allocation set value is output to the communication interface of each adjustable equipment cabinet and energy storage tank to generate multi-energy complementary heating control command.

8. A multi-energy complementary intelligent control system for solar heating, characterized in that, include: The calculation module is used to collect multi-source operation data of photovoltaic system and heating network to construct time series input matrix. Based on irradiance, cloud cover and heat load change rate, the time series input matrix is ​​segmented into states, and the irradiance slope, temperature difference accumulation, load gradient and state reversal index of each segment are calculated. The generation module is used to input each segment into the Long Short-Term Memory network to calculate the base value of the forget gate based on the current input and the previous hidden state. When the cloud cover changes abruptly, the base value of the forget gate is adjusted based on the continuous prediction residual characteristics to obtain the updated forget coefficient. The updated forget coefficient and the previous unit state are jointly weighted and aligned in dimension according to the dimensionless irradiance slope, temperature difference accumulation and load gradient to obtain the state retention amount. The memory compression factor is generated according to the state reversal index. The state writing amount is obtained by constraining the input gate output and candidate states. The state retention amount and the writing amount are superimposed to update the unit state and generate the solar power output sequence and fluctuation index. The adjustment module is used to construct an objective function based on the solar power output sequence, fluctuation index, heat load forecast, thermal storage status, and electricity price sequence. It solves the objective function and constraint equation set to obtain the time-period allocation of photovoltaic-driven heat pumps, gas boilers, thermal storage charging and discharging, and water supply temperature. It calculates the prediction reliability of the solar power output sequence, adjusts the heat pump absorption and load gap thresholds according to the prediction reliability, and generates multi-energy complementary heating control instructions in combination with the time-period allocation.

9. The system according to claim 8, characterized in that, The process of collecting multi-source operational data from photovoltaic systems and heating networks to construct a time-series input matrix includes: Historical discrete time series of photovoltaic array irradiance, module temperature, ambient temperature, cloud cover, heating network supply and return water temperature, heat load, electricity price, and state of charge of thermal storage units are acquired using data acquisition sensors. For time series with inconsistent frequencies and misalignments, the arithmetic mean interpolation of adjacent time points is used to align the data nodes and integrate the time series into a unified sampling frequency with a consistent time interval. The aligned historical data points are then stacked and merged as independent input vectors in the time domain to generate a two-dimensional time series input matrix that expresses the complete temporal development characteristics of multi-source data.

10. The system according to claim 8, characterized in that, The process of segmenting the time-series input matrix based on irradiance, cloud cover, and heat load change rate, and calculating the irradiance slope, cumulative temperature difference, load gradient, and state reversal index for each segment, includes: The time-series input matrix is ​​processed using a time sliding window to calculate the difference between parameters corresponding to adjacent time nodes in the time dimension, resulting in the irradiance change rate, cloud cover change rate, and heat load change rate. Waveform regions where the absolute value of the change rate continuously exceeds a set segment constant are extracted as unstable abrupt change segments. The remaining main data is truncated at a set fixed interval to form relatively static segments, thus constituting independent state segments. The irradiance slope is obtained by dividing the irradiance difference at the start and end positions of the state segment by the time span of the segment. The temperature difference accumulation is obtained by summing the absolute values ​​of the component temperature and ambient temperature differences at each moment within the state segment. The load gradient is obtained by dividing the heat power difference at the start and end positions of the state segment by the time span of the segment. The state reversal index is obtained by counting the cumulative number of sign reversals of the time-domain waveform trends of parameters within the state segment, from rising to falling or from falling to rising.

Images