MSWI air volume control method and system based on bionic hybrid model and dynamic modeling

Abstract

The present disclosure provides a MSWI air volume control method and system based on a bionic hybrid model and dynamic modeling, which relates to the technical field of MSWI air volume prediction and control, including: obtaining continuous key operating condition signals of the MSWI system, pre-processing them and converting them into a pulse sequence; inputting the pulse sequence into an optimized bionic hybrid neural network model, introducing a short-term plasticity mechanism and a long-term plasticity mechanism, and predicting and outputting the MSWI key air volume at a future moment; dividing the MSWI incineration process into multiple areas, establishing corresponding material balance, energy balance and reaction rate mathematical models for each area, simulating the physical and chemical reaction processes of waste incineration, and obtaining a simulated theoretical air volume; fusing the predicted MSWI key air volume with the simulated theoretical air volume, obtaining a comprehensive prediction value through a weighted method, and using the comprehensive prediction value to perform integrated control of the MSWI air volume.

Description

Technical Field

[0001] The present disclosure relates to the technical field of MSWI air volume prediction and control, and in particular to an MSWI air volume control method and system based on a bionic hybrid model and dynamic modeling. Background Art

[0002] The statements in this section merely provide background information related to the present disclosure and do not necessarily constitute prior art.

[0003] Municipal solid waste incineration (MSWI) has become a mainstream technology for waste treatment due to its efficient volume reduction and energy recovery capabilities. In MSWI projects, the furnace air volume (primary air, secondary air, and excess air coefficient) determines combustion stability, thermal efficiency, and pollutant emissions.

[0004] On the one hand, the complex and variable composition of solid waste (influenced by residents' living habits and the prevalence of waste sorting) leads to large fluctuations in the calorific value of the materials entering the furnace, causing problems such as unstable combustion, coking in the furnace, and ash accumulation. Traditional control methods rely on the experience of operating experts to adjust air volume, but manual judgment is subject to variability and physical limitations, making continuous optimization difficult.

[0005] In addition, traditional power plants generally use PID or empirical tables plus PLC sequential control: operators manually adjust the damper based on single-point signals such as flue gas temperature and furnace negative pressure. However, the composition of MSW fluctuates greatly with season, region, and time of entry into the furnace. The combustion process has strong coupling, long time delay, and nonlinear characteristics. Linear PID alone cannot simultaneously suppress furnace temperature fluctuations and NO. x If the standard is exceeded, the alternating cycles of "excess air volume → heat loss" or "insufficient oxygen → black smoke" will often occur. Summary of the Invention

[0006] In order to solve the above problems, the present disclosure proposes an MSWI air volume control method and system based on a bionic hybrid model and dynamic modeling. The hybrid bionic neural network (Bio-HybridNet) that integrates the bionic neural network and the synaptic plasticity mechanism is optimized by a bionic optimization algorithm. The air volume prediction results of the hybrid bionic neural network are fused with the simulation results of the dynamic model of the MSWI incineration process. By dynamically adjusting the grate speed, excess air coefficient and air volume proportional coefficient, the control distribution of the primary air volume and the secondary air volume is realized to meet the needs of actual operation.

[0007] According to some embodiments, the present disclosure adopts the following technical solutions:

[0008] The MSWI air volume control method based on bionic hybrid model and dynamic modeling includes:

[0009] Obtain the continuous key operating condition signals of the MSWI system, and convert them into pulse sequences after preprocessing;

[0010] The pulse sequence is input into the optimized biomimetic hybrid neural network model. The pulse signal vector of the pulse sequence is first extracted through the pulse neural network module. The pulse signal vector is then transferred to the LSTM module. During the transfer process, a short-term plasticity mechanism is introduced. By introducing a dynamic weight attenuation coefficient and recovery coefficient, the response of synaptic transmission is adjusted to the sudden input. A long-term plasticity mechanism is also introduced between the input connection and the gate control unit within the LSTM module. The network weight is persistently adjusted for positive and negative feedback information respectively. Finally, the LSTM module predicts and outputs the MSWI critical air volume at the future moment.

[0011] The MSWI incineration process is divided into multiple zones, and corresponding material balance, energy balance and reaction rate mathematical models are established for each zone to simulate the physical and chemical reaction processes of waste incineration and obtain the simulated theoretical air volume.

[0012] The predicted MSWI critical air volume is fused with the simulated theoretical air volume, and a comprehensive prediction value is obtained through a weighted method. The comprehensive prediction value is used for integrated control of MSWI air volume.

[0013] According to some embodiments, the present disclosure adopts the following technical solutions:

[0014] The MSWI air volume control system based on bionic hybrid model and dynamic modeling includes:

[0015] The signal acquisition module is used to obtain the continuous key working condition signals of the MSWI system and convert them into pulse sequences after preprocessing;

[0016] The feature extraction and air volume prediction module is used to input the pulse sequence into the optimized biomimetic hybrid neural network model. The pulse signal vector of the pulse sequence is first extracted through the pulse neural network module, and then the pulse signal vector is transmitted to the LSTM module. During the transmission process, a short-term plasticity mechanism is introduced. By introducing a dynamic weight attenuation coefficient and recovery coefficient, the response of synaptic transmission is adjusted to the sudden input. A long-term plasticity mechanism is introduced between the input connection and the gate control unit within the LSTM module. The network weight is persistently adjusted for positive and negative feedback information. Finally, the LSTM module predicts and outputs the MSWI critical air volume at the future moment.

[0017] The dynamic simulation module is used to divide the MSWI incineration process into multiple zones, establish corresponding material balance, energy balance and reaction rate mathematical models for each zone, simulate the physical and chemical reaction process of waste incineration, and obtain the simulated theoretical air volume;

[0018] The integrated control module is used to fuse the predicted MSWI key air volume with the simulated theoretical air volume, obtain the comprehensive prediction value through the weighted method, and use the comprehensive prediction value to perform MSWI air volume integrated control.

[0019] According to some embodiments, the present disclosure adopts the following technical solutions:

[0020] A computer program product includes a computer program. When the computer program is executed by a processor, the computer program implements the MSWI air volume control method based on a bionic hybrid model and dynamic modeling.

[0021] According to some embodiments, the present disclosure adopts the following technical solutions:

[0022] A non-transitory computer-readable storage medium is used to store computer instructions. When the computer instructions are executed by a processor, the MSWI air volume control method based on a bionic hybrid model and dynamic modeling is implemented.

[0023] According to some embodiments, the present disclosure adopts the following technical solutions:

[0024] An electronic device includes: a processor, a memory, and a computer program; wherein the processor is connected to the memory, and the computer program is stored in the memory. When the electronic device is running, the processor executes the computer program stored in the memory to enable the electronic device to implement the MSWI air volume control method based on a bionic hybrid model and dynamic modeling.

[0025] Compared with the prior art, the present invention has the following beneficial effects:

[0026] The disclosed MSWI air volume control method based on a bionic hybrid model and dynamic modeling uses a bionic optimization algorithm to optimize a hybrid bionic neural network that integrates a bionic neural network and a synaptic plasticity mechanism. The bionic hybrid neural network captures the characteristics of the data set, synaptic plasticity enhances the rapid adaptive ability of the network, and the bionic optimization algorithm ensures the global optimal parameters, thereby achieving efficient, stable, and robust prediction of the MSWI air volume. At the same time, the incineration process of the MSWI is dynamically modeled to accurately simulate the physical and chemical reaction processes of waste incineration. The two are integrated to achieve integrated air volume control. By dynamically adjusting the grate speed, excess air coefficient, and air volume proportional coefficient, a reasonable distribution of primary and secondary air volumes is achieved to meet the needs of actual operation.

[0027] The disclosed MSWI air volume control method, based on a biomimetic hybrid model and dynamic modeling, simulates the learning ability of a biological brain by introducing two levels of synaptic plasticity: short-term plasticity (STP) and long-term plasticity (LTP / LTD) between the SNN and LSTM, and within the LSTM. STP simulates the rapid consumption and recovery of neurotransmitters between synapses, enabling the network to respond more sensitively to sudden, transient input changes. LTP (long-term potentiation) and LTD (long-term depression) are used for persistent, global learning and memory of network weights, primarily implemented in the gating and input connections within the LSTM.

[0028] The MSWI air volume control method based on the bionic hybrid model and dynamic modeling disclosed in the present invention divides the MSWI incineration process into multiple zones, establishes corresponding material balance, energy balance and reaction rate mathematical models for each zone, simulates the physical and chemical reaction processes of waste incineration, and obtains the simulated theoretical air volume;

[0029] This disclosed MSWI air volume control method, based on a biomimetic hybrid model and dynamic modeling, dynamically models the MSWI incineration process, accurately simulating the physical and chemical reactions involved. By combining physical and chemical models with machine learning models, it forms a more accurate prediction and optimization framework. The integrated control strategy dynamically adjusts grate speed, excess air coefficient, and air volume proportionality coefficient to bring predicted values ​​close to theoretical calculated values, thereby achieving a reasonable distribution of primary and secondary air volumes to meet actual operational requirements. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] The accompanying drawings, which constitute a part of the present disclosure, are used to provide a further understanding of the present disclosure. The exemplary embodiments of the present disclosure and their descriptions are used to explain the present disclosure and do not constitute an improper limitation to the present disclosure.

[0031] Figure 1 Flowchart of the MSWI air volume control method based on the bionic hybrid model and dynamic modeling according to an embodiment of the present disclosure;

[0032] Figure 2 Schematic diagram of the LSTM structure of an embodiment of the present disclosure;

[0033] Figure 3 Schematic diagram of the modeling of the MSWI incineration process according to an embodiment of the present disclosure;

[0034] Figure 4 is an integrated control flow chart of an embodiment of the present disclosure;

[0035] Figure 5Comparison of primary air volume prediction results of different prediction models according to the embodiments of the present disclosure;

[0036] Figure 6 Comparison of secondary air volume prediction results of different prediction models according to the embodiments of the present disclosure;

[0037] Figure 7 A comparison between the actual furnace temperature value and the output of the integrated model in the embodiment of the present disclosure;

[0038] Figure 8 A comparison between the actual value of the main steam flow rate and the output of the integrated model in the embodiment of the present disclosure;

[0039] Figure 9 A comparison between the actual value of flue gas oxygen content and the output of the integrated model according to an embodiment of the present disclosure;

[0040] Figure 10 Comparison between the actual value of the primary air volume and the model predicted value in the embodiment of the present disclosure;

[0041] Figure 11 Comparison between the actual value of the secondary air volume and the model predicted value in the embodiment of the present disclosure. DETAILED DESCRIPTION

[0042] The present disclosure will be further described below with reference to the accompanying drawings and embodiments.

[0043] It should be noted that the following detailed descriptions are illustrative and intended to provide further explanation of the present disclosure. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which the present disclosure belongs.

[0044] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present disclosure. As used herein, unless the context clearly indicates otherwise, the singular form is intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, they indicate the presence of features, steps, operations, devices, components and / or combinations thereof.

[0045] Example 1

[0046] In one embodiment of the present disclosure, a method for controlling MSWI air volume based on a bionic hybrid model and dynamic modeling is provided, comprising:

[0047] Step 1: Obtain the continuous key operating condition signal of the MSWI system, pre-process it and convert it into a pulse sequence;

[0048] Step 2: Input the pulse sequence into the optimized biomimetic hybrid neural network model. First, the pulse signal vector of the pulse sequence is extracted through the pulse neural network module. Then, the pulse signal vector is transferred to the LSTM module. During the transfer process, a short-term plasticity mechanism is introduced. By introducing the dynamic attenuation coefficient and recovery coefficient of the weight, the response of synaptic transmission is adjusted to the sudden input. A long-term plasticity mechanism is introduced between the input connection and the gate control unit within the LSTM module. The network weight is persistently adjusted for positive and negative feedback information respectively. Finally, the LSTM module predicts and outputs the MSWI critical air volume at the future moment.

[0049] Step 3: Divide the MSWI incineration process into multiple zones, establish corresponding material balance, energy balance, and reaction rate mathematical models for each zone, simulate the physical and chemical reaction processes of waste incineration, and obtain the simulated theoretical air volume;

[0050] Step 4: Fuse the predicted MSWI critical air volume with the simulated theoretical air volume, obtain the comprehensive prediction value through the weighted method, and use the comprehensive prediction value to perform integrated control of the MSWI air volume.

[0051] As an embodiment, the disclosed MSWI air volume control method based on bionic hybrid model and dynamic modeling optimizes the hybrid bionic neural network through a bionic optimization algorithm. The bionic hybrid neural network (Bio-HybridNet) includes a pulse neural network module, an LSTM module and a synaptic plasticity module. Synaptic plasticity enhances the rapid adaptive ability of the network to achieve efficient, stable and robust prediction of the MSWI air volume. At the same time, the incineration process of MSWI is dynamically modeled to accurately simulate the physical and chemical reaction processes of waste incineration. Combining the two, a new type of air volume integrated control method is constructed. By dynamically adjusting the grate speed, excess air coefficient and air volume proportional coefficient, a reasonable distribution of primary air volume and secondary air volume is achieved to meet the needs of actual operation. As Figure 1 The specific implementation process is as follows:

[0052] Step 1: Obtain the continuous key operating condition signal of the MSWI system, pre-process it and convert it into a pulse sequence;

[0053] Specifically, continuous key operating condition signals are collected from the MSWI system, including air volume, temperature, furnace temperature, air inlet pressure, flue gas CO concentration, and flue gas NO X Concentration data, etc.

[0054] Afterwards, the collected key operating condition signals are preprocessed, including denoising and normalization, to ensure that the operating condition signals have a stable numerical range and a high signal-to-noise ratio in the subsequent processing process.

[0055] Furthermore, the pre-processed continuous working condition signal is converted into a pulse sequence using frequency coding. The corresponding pulse frequency is set according to the signal amplitude - a higher frequency pulse is generated when the signal value is large, and a lower frequency pulse is generated when the value is small, thereby mapping the continuous working condition signal encoding into a discrete pulse sequence pattern:

[0056]

[0057] in, S ( t ) is the time t The original signal, S min and S max are the minimum and maximum values ​​of the working condition signal respectively, f min and f max These are the minimum and maximum pulse frequencies that are set. In this way, when the signal value is large, a higher pulse output frequency is generated, and vice versa, a lower frequency pulse is output.

[0058] Step 2: Input the pulse sequence into the optimized biomimetic hybrid neural network model. First, the pulse signal vector of the pulse sequence is extracted through the pulse neural network module. Then, the pulse signal vector is passed to the LSTM module. During the transmission process, a short-term plasticity mechanism is introduced. By introducing the dynamic attenuation coefficient and recovery coefficient of the weight, the response of synaptic transmission is adjusted to the sudden input. A long-term plasticity mechanism is introduced between the input connection and the gate control unit within the LSTM module. The network weight is persistently adjusted for positive and negative feedback information respectively. Finally, the LSTM module predicts and outputs the MSWI critical air volume at the future moment. The details are as follows:

[0059] Step 21: Structure and optimization process of the bionic hybrid neural network model;

[0060] First, the structure of the bionic hybrid neural network model is as follows:

[0061] The bionic hybrid neural network model includes a pulse neural network module, an LSTM module and a synaptic plasticity module. The pulse neural network module includes pulse neurons and multi-layer SNN convolutions. The pulse neurons are designed based on the Leaky Integrate-and-Fire (LIF) model. The LIF model describes the accumulation, leakage and discharge process of neuronal membrane potential after reaching the threshold. Each layer of the multi-layer SNN convolution adopts a convolution kernel and delay window with bionic settings to simulate the response of the biological visual cortex to signals of different frequencies. A time domain pooling layer is set after each convolution layer to extract local time features by performing local statistics on the pulse signal (such as pulse counting or membrane potential averaging), reducing the data dimension while retaining key dynamic characteristics.

[0062] The LSTM module is a long short-term memory (LSTM) network, a special type of recurrent neural network (RNN) consisting of multiple layers of LSTM. Each layer contains several LSTM neurons. The first layer is used to capture preliminary temporal dependencies, and subsequent layers are used to extract deeper time series. The input gate, forget gate, and output gate of the LSTM unit are configured to ensure that the network can dynamically control state updates and information transfer. The LSTM module disclosed herein includes at least two layers of LSTM networks.

[0063] The synaptic plasticity module includes short-term plasticity mechanism and long-term plasticity mechanism. The short-term plasticity mechanism is introduced between the multi-layer SNN convolution and the LSTM module, and the long-term plasticity mechanism is introduced inside the LSTM module. The short-term plasticity mechanism simulates the rapid consumption and recovery of neurotransmitters between synapses, introduces a short-term weight adjustment factor, and modulates signal transmission, so that the network shows higher sensitivity to sudden and instantaneous input changes; the long-term plasticity mechanism adopts a weight correction algorithm based on error feedback, so that when the network prediction error is large, the relevant synaptic weights are corrected multiple times to achieve long-term enhancement or inhibition, so as to improve the memory effect of the overall model.

[0064] As an embodiment, the optimization training process of the bionic hybrid neural network model is as follows:

[0065] The bionic hybrid neural network model uses a combination of online and offline optimization. During the initial training phase, global optimization (offline) is performed based on historical data. During operation, a periodic, triggered optimization method is used. When the error exceeds a threshold, the bionic optimization algorithm is called to adjust the hyperparameters.

[0066] Step 1: Determine the hyperparameter optimization goal

[0067] Identify the hyperparameters that need to be tuned: including the pulse threshold and membrane time constant of the SNN, the number of neurons of the LSTM network, the learning rate and other key parameters.

[0068] Set the objective function and use the prediction error (such as mean square error MSE) as the optimization indicator, with the goal of minimizing the prediction error.

[0069] The objective function can be expressed as:

[0070]

[0071] in, y i is the actual air volume data; i is the wind volume data predicted based on the current hyperparameters; N is the sample size. ( θ SNN , τ m , n LSTM , η ) represents a set of hyperparameters.

[0072] Step 2: Select the improved bionic optimization algorithm

[0073] This algorithm uses an intelligent optimization algorithm with biomimetic characteristics. Its update mechanism is based on the concept of swarm search and is designed to solve multi-modal optimization problems. The algorithm is improved by introducing a dynamic adjustment convergence factor to enhance local search capabilities and adapt to complex optimization spaces. The specific process includes:

[0074] Set the candidate solution set { x i ( t )} i=1 M (Each candidate solution x i represents a set of hyperparameters [ θ SNN , τ m , n LSTM , η ]), and determine the current optimal solution x best ( t ).

[0075] The general formula for updating candidate solutions is:

[0076]

[0077] in, α 、 β 、 γ is the adjustment coefficient, x rand (t ) is a set of parameters randomly selected from the candidate solutions, t Indicates the number of iterations.

[0078] The update rule aims to make the candidate solutions move closer to the optimal solution while retaining a certain degree of randomness to escape from the local optimal solution.

[0079] Step 3: Global optimization of parameters in the offline phase

[0080] The bionic optimization algorithm was trained offline using historical MSWI operation data, and a set of preliminary optimal hyperparameters was obtained through global search. After the initial parameters were determined, they were used as the default values ​​when the system started.

[0081] During the offline optimization process, each candidate solution is continuously updated using the following formula: x i Until the global optimal solution converges:

[0082]

[0083] Finally, the global optimal hyperparameters As the default parameter when the system starts.

[0084] Step 4: Online error monitoring and triggering mechanism

[0085] During the real-time operation of the system, the error between the predicted result and the actual value is calculated in real time E ( t ), where the error calculation formula is:

[0086]

[0087] in: N t is the number of samples in the current time window; y i ( t )and i ( t ) are the actual value and the predicted value respectively.

[0088] when E ( t ) exceeds the preset threshold E th When E ( t )> E th , triggering the online parameter adjustment mechanism and calling the bionic intelligent algorithm for local search and parameter adjustment.

[0089] When the error exceeds the preset threshold, the online parameter adjustment mechanism is triggered, and the improved bionic optimization algorithm is called to perform local search and parameter adjustment.

[0090] Step 5: Parameter update and verification

[0091] The new parameters obtained by online bionic intelligent algorithm optimization are updated to the Bio-HybridNet model. The new parameters are recorded as The effect of parameter adjustment is verified through a feedback closed loop. If the prediction error is reduced, the new parameters are maintained; otherwise, the optimization process is continued until the set error target is reached.

[0092] If the updated mean square error satisfies:

[0093]

[0094] in, MSE new Indicates the use of new hyperparameters x new The mean square error of the posterior model, MSE prev represents the mean squared error of the model before parameter tuning. If the updated mean squared error satisfies the above formula, the new parameters are saved; otherwise, local optimization continues until the model error reaches the target. This process forms a closed-loop feedback system, ensuring the continuous effectiveness of online tuning, allowing the system to dynamically adapt to changing operating conditions.

[0095] The optimized and trained bionic hybrid neural network is used in MSWI wind volume prediction.

[0096] Step 22: Input the pulse sequence into the optimized bionic hybrid neural network model, first entering the pulse neural network module. The specific data processing process of the pulse neural network module is as follows:

[0097] First, after the pulse sequence is input into the biomimetic hybrid neural network model, it first enters the spiking neural network module, which first passes through the spiking neurons designed based on the Leaky Integrate-and-Fire (LIF) model. The LIF model describes the accumulation, leakage, and discharge process of the neuron membrane potential when the threshold is reached; key parameters are set, including the pulse threshold, membrane time constant, and reset potential, to ensure that the neuron can accurately respond to the dynamic changes in the input pulse sequence. Using the Leaky Integrate-and-Fire (LIF) model to describe the membrane potential changes of spiking neurons, its dynamic process can be described as:

[0098]

[0099] in, V ( t) indicates the time t membrane potential; τ m is the membrane time constant; V rest is the resting potential; R is the resistance of the neuron; I ( t ) is the input current (which can be composed of an input pulse sequence after pulse frequency encoding).

[0100] When the membrane potential V ( t ) reaches or exceeds the set threshold V th When the neuron generates a pulse, the output signal O ( t ) is expressed as:

[0101]

[0102] After the pulse is generated, the membrane potential is reset to V reset .

[0103] Furthermore, the pulse output signal generated by the neuron is input into the multi-layer SNN convolution. In the multi-layer SNN convolution network, each layer adopts a convolution kernel and a time delay window with a biomimetic setting to simulate the response of the biological visual cortex to signals of different frequencies. A time domain pooling layer is set after each convolution layer to extract local time features by performing local statistics on the pulse signal (such as pulse counting or membrane potential averaging), thereby reducing the data dimension while retaining key dynamic characteristics. In the multi-layer pulse convolution layer, the output of each layer can be regarded as a discrete convolution operation of the input pulse sequence and the convolution kernel. If the input pulse sequence is recorded as O ( t ), the convolution kernel is w ( τ ), then the convolution output Y ( t ) can be expressed as:

[0104]

[0105] in: T is the time length of the convolution kernel, w ( τ ) is the convolution kernel in the delay τ This convolution operation can effectively extract the local time features in the input pulse sequence.

[0106] In order to reduce the data dimension and retain key dynamic features, pooling operations (such as average pooling) are often used. The output pulse sequence obtained after multi-layer pulse convolution and pooling is converted into a fixed-dimensional feature vector. This feature vector will serve as the input of the downstream long short-term memory network (LSTM) time series modeling module to further capture the time series dependencies and long-term patterns in the working condition signal. Assume that there are N pulse outputs in a small time window, and the pulse signal vector of the pooled output is P It can be calculated as:

[0107]

[0108] The formula represents the pulse count or membrane potential average within a time window, which can be used as the input feature of the downstream LSTM module.

[0109] As an embodiment, the spiking neural network module is tested offline to verify the effectiveness of the encoding and extraction process by comparing the characteristic distribution and key dynamic information of the signal before and after conversion;

[0110] According to the test results and actual working condition feedback, the pulse coding strategy, neuron parameters and convolution layer settings are adjusted to ensure that the module output characteristics can accurately reflect the instantaneous and dynamic change characteristics of the signal in the MSWI system.

[0111] Step 23: The extracted pulse signal vector of the pulse train is passed to the LSTM module for further processing. To simulate the learning ability of the biological brain, a two-level synaptic plasticity mechanism is introduced between the multi-layer SNN convolution and the LSTM, and within the LSTM: short-term plasticity (STP) and long-term plasticity (LTP / LTD). Short-term plasticity (STP) is introduced during the transmission process between the multi-layer SNN convolution and the LSTM module. STP simulates the rapid consumption and recovery of neurotransmitters at synapses, enabling the network to respond more sensitively to sudden and transient input changes. The details are as follows:

[0112] A short-term plasticity mechanism is established between SNN and LSTM to simulate the rapid consumption and recovery of neurotransmitters between synapses. The commonly used description model is the Tsodyks-Markram model, whose basic formula is as follows:

[0113] Available resources change dynamically:

[0114]

[0115] in: R (t ) indicates the time t The proportion of synaptic resources available when ; τ rec is the time constant for resource recovery; u ( t ) is the utilization rate, which indicates the proportion of synaptic resources used when a pulse is triggered; is the pulse trigger function, when the neuron is t spike It is 1 when a pulse is generated, otherwise it is 0.

[0116] Utilization dynamic update:

[0117]

[0118] in: U is the initial utilization constant; τ facil is the time constant for utilization recovery or decay.

[0119] At each spike, the effective response of synaptic transmission (i.e., the short-term weight adjustment factor) can be expressed as:

[0120]

[0121] This factor is used to modulate signal transmission, making the network more sensitive to sudden and instantaneous input changes.

[0122] As an example, the short-term plasticity mechanism works by introducing dynamic weight attenuation and recovery coefficients during signal transmission, allowing synaptic transmission to rapidly adjust to sudden input changes. This can increase the network's sensitivity to transient and drastically changing inputs, enabling the model to quickly respond to transient fluctuations.

[0123] Furthermore, the pulse signal vector converted by the pulse neural network module is used as the input of the LSTM module. The pulse signal vector extracted by the pulse neural network module is recorded as the input sequence { x 1, x 2,…, x T},in x t ∈R d Indicates at time t ; each input feature vector is normalized, for example, using zero mean and unit variance normalization, and the calculation formula is:

[0124]

[0125] in, is the normalized output feature, μ is the mean of the input data, σ is the standard deviation.

[0126] The LSTM module of this disclosure is designed with at least two layers of LSTM networks, each layer containing several LSTM neurons. The first layer is used to capture preliminary temporal dependencies, and subsequent layers are used to further abstract features and extract deep time series patterns.

[0127] like Figure 2 As shown in the figure, configure the input gate, forget gate and output gate of the LSTM unit to ensure that the network can dynamically control the state update and information transmission. l Tier t The hidden state and cell state of each time step are recorded as and , the corresponding input vector is recorded as , then the gated machine state update process is as follows: the retention and update ratio of the cell state is determined by the forget gate, input gate and candidate state; Based on the fusion of new information generation The current hidden state is output and used as the input for the next layer or time step. This gating mechanism ensures that the network can preserve long-range dependencies while quickly incorporating new information when processing non-stationary, multi-scale sequences.

[0128] Among them, the forget gate:

[0129]

[0130] Input Gate:

[0131]

[0132] Candidate cell states:

[0133]

[0134] Cell status update:

[0135]

[0136] Output gate:

[0137]

[0138] Hidden state output:

[0139]

[0140] Where σ(•) represents the Sigmoid function; is element-wise multiplication; Wf (l) 、 W i (l) 、 W C (l) and W o (l) is the weight matrix, b f (l) 、 b i (l) 、 b C (l) and b o (l) is the bias vector; for the first layer, x t (1) is the feature vector normalized in Step 1; for subsequent layers, x t (l) The output of the previous layer h t (l-1) .

[0141] In a specific embodiment, a long-term plasticity (LTP / LTD) mechanism is introduced within the LSTM. LTP (long-term potentiation) and LTD (long-term depression) are used to perform persistent, global learning and memory of network weights, mainly implemented in the gating and input connections within the LSTM.

[0142] Long-term plasticity regulation strategies are introduced in key connection layers within LSTM (such as between input connections and gating units). LTP (long-term potentiation) and LTD (long-term depression) adjust network weights persistently based on positive and negative feedback information, respectively. The details are as follows:

[0143] A weight correction algorithm based on error feedback is adopted, so that when the network prediction error is large, the relevant synaptic weights can be corrected multiple times to achieve long-term enhancement (or inhibition) to improve the memory effect of the overall model.

[0144] Among them, the weight correction formula based on error feedback is:

[0145]

[0146] Among them, Δ w ij Represents connected neurons i and jThe weight update of η Learning rate; x i is the output of the neurons in the front layer; δ j is the error signal of the neurons in the latter layer (which can be calculated by back propagation), representing positive feedback (for LTP) or negative feedback (for LTD).

[0147] Furthermore, the weight update formula (with memory effect) is:

[0148]

[0149] Through multiple iterations, when the prediction error is large, the relevant weights are updated multiple times in a positive direction to achieve long-term potentiation (LTP); conversely, if the error prompt is too strong, long-term depression (LTD) is achieved through negative updates.

[0150] The short-term and long-term plasticity mechanisms are used in parallel in the network. Short-term plasticity is responsible for the rapid adaptation of the network, while long-term plasticity ensures the long-term stable learning effect of the network.

[0151] In each training cycle, the attenuation / recovery parameters of STP and the weight update amplitude of LTP / LTD are dynamically adjusted according to the prediction error and real-time feedback information to ensure that the two mechanisms work in coordination.

[0152] In each signal transmission process, the weight is modulated for a short time and then combined with the long-term update, and the comprehensive weight is recorded as W ij ∗ ( t ), which can be expressed as:

[0153]

[0154] in, w ij ( t ) is the weight adjusted by long-term plasticity; f STP ( t )= u ( t ) R ( t ) is the short-term plasticity modulation factor.

[0155] As an embodiment, in each training cycle, updates are performed simultaneously, specifically as follows:

[0156] For the STP module: Using the aforementioned Tsodyks-Markram model formula, update u ( t )and R( t );

[0157] For LTP / LTD modules: Based on error feedback δ j , using the weight update formula to correct w ij ( t ).

[0158] At the same time, a joint adjustment strategy can be set, such as normalizing the update step size during the update process to ensure that the gradients of short-term modulation and long-term memory have a reasonable numerical ratio.

[0159] As an example, a preliminary set of STP and LTP / LTD parameters was obtained through experiments using initial offline data. During system operation, the error between network output and actual air volume was monitored in real time. When the error exceeded the set range, an online tuning mechanism was invoked to correct the STP and LTP / LTD parameters to ensure that the system can continuously adapt to changing operating conditions.

[0160] Verification and feedback: The network with enhanced synaptic plasticity mechanisms is tested both offline and online, comparing its prediction performance to that without the plasticity mechanism. This ensures that the network's predictions respond faster and with lower error when responding to abnormal or sudden signals, thereby improving overall prediction performance and robustness. Parameters are adjusted based on the verification results, and the final mechanism is embedded in the system, forming a closed-loop adaptive prediction and control solution.

[0161] As an example, key hyperparameters of the LSTM network are determined, including the number of hidden layer neurons, time step, batch size, and learning rate. The initial values ​​of these parameters can be obtained by fitting offline historical data and then further optimized through online tuning. The training and learning process includes:

[0162] Historical MSWI operation data is used for offline training, the mean square error (MSE) is used as the loss function, and the backpropagation through time (BPTT) algorithm is used to update parameters.

[0163] An early stopping strategy is implemented during training to prevent overfitting and ensure the generalization ability of the model.

[0164] As an example, a trained LSTM model is used to predict real-time data, outputting the MSWI critical air volume for the next moment or a period in the future. The prediction results serve as the core basis for the feedback control system and are transmitted to the integrated control module for subsequent regulation.

[0165] As an embodiment, the process constitutes a closed-loop feedback system to ensure that the online tuning is continuously effective, thereby enabling the system to dynamically adapt to changes in operating conditions. Figure 5 and Figure 6 The prediction results for primary and secondary air volume using different models are shown in Table 1. Table 1 compares the air volume prediction results using different prediction models. It can be seen that the model using the optimization algorithm significantly improves the prediction accuracy of air volume, reducing the RMSE for the primary and secondary air volume prediction tasks by 2.21% and 2.00%, respectively.

[0166] Table 1 Comparison of wind volume prediction results of different prediction models

[0167]

[0168] Step 3: Divide the MSWI incineration process into multiple zones, establish corresponding material balance, energy balance, and reaction rate mathematical models for each zone, simulate the physical and chemical reaction processes of waste incineration, and obtain the simulated theoretical air volume;

[0169] Specifically, step 31: incineration area division and model establishment

[0170] First, the physical parameters of MSW and furnace are collected to facilitate dynamic modeling research. The parameters are shown in Table 2.

[0171] Table 2 Parameters of MSWI

[0172]

[0173] like Figure 3 As shown in the figure, the MSWI incineration process is divided into seven areas: hopper area (A), feeder area (B), drying area (C), solid phase combustion area (D), solid phase ash area (E), gas phase combustion area (F) and waste heat boiler area (G).

[0174] The material transfer, heat exchange and chemical reaction processes in each area are modeled, and the corresponding material balance, energy balance and reaction rate mathematical models are established respectively.

[0175] Among them, (1) Material balance:

[0176] For a certain area i (e.g. drying area), the material balance can be expressed as:

[0177]

[0178] in, m i Indicates area i The instantaneous mass of the material in m in,i andm out,i Represent the feed and discharge flow rates of the area respectively.

[0179] Step 32: Establish basic equations for physical and chemical reactions

[0180] Based on experimental data and theoretical foundations, mathematical formulas for material flow and chemical reactions in each area are compiled, including: calculation of the distribution of primary and secondary air in each area, thermal radiation heat transfer, convection heat transfer, and chemical reaction conditions (such as the critical parameters of the reaction between volatile substances and fixed carbon).

[0181] (1) Air distribution calculation

[0182] In an incinerator, the primary air's main function is to provide the oxygen required for combustion. Primary air enters the combustion chamber from the bottom, ensuring that the MSW is fully exposed to oxygen and reacts completely. Secondary air mainly provides additional oxygen to the combustion chamber, making combustion more thorough, thereby reducing black smoke and harmful emissions. Considering the distribution of primary and secondary air in each area, the primary air volume is set to u I , i and the secondary air volume is u II,i The following calculation formula can be used:

[0183]

[0184] At the same time, the allocation ratio is set according to the needs of different regions:

[0185]

[0186]

[0187] in, α i For the region i Proportional coefficient of internal primary air.

[0188] (2) Chemical reaction rate

[0189] For the combustion reaction of volatile substances and fixed carbon, the reaction rate can be described by the Arrhenius type formula r :

[0190]

[0191] in, k 0 is the pre-exponential factor, E a is the reaction activation energy, R is the gas constant, T i For the regioni temperature, C is the reactant concentration or the mass ratio of related materials.

[0192] (3) Thermal balance, mainly the heat transfer between radiation and convection for the region i The thermal balance of assuming that there are two main heat transfer processes, convection and radiation, can be expressed as:

[0193]

[0194]

[0195]

[0196] in, is the convective heat transfer; It is radiant heat transfer; Indicates area i Other heat losses; c p is the specific heat capacity of the material; T air , i For the region i The temperature of the air inside, T flame is the temperature of the flame in the combustion area; T i For the region i The temperature inside, m i Indicates area i The instantaneous mass of the material in h i 、 A i 、 are the corresponding convective heat transfer coefficient, heated area and emissivity respectively; σ is the Stefan-Boltzmann constant, which is 5.67×10 -8 W / (m 2 ·K 4 ). The convective heat transfer coefficient may depend on the flow state and geometric dimensions; in radiation heat transfer, and A i It can be obtained by fitting experimental data.

[0197] Step 33: Energy balance model construction

[0198] Taking the entire incinerator as a whole, an energy balance equation is established based on parameters such as the lower calorific value of MSW, inlet and outlet air temperature, and combustion heat release, and the heat flowing into the incinerator is analyzed:

[0199]

[0200]

[0201] in, Q in and Q out are the energy entering and leaving the incinerator respectively; m air for air quality; c p,air is the specific heat capacity of air; T in and T out are the inlet and outlet temperatures, respectively; m A,in It represents the amount of MSW entering the incinerator; c p,air is the specific heat capacity of air; q MSW It is the basic low heat of 1kg MSW; T in,I ，T in,II ，T out,I and T out,II are the inlet and outlet temperatures of the two fans, c in,I 、c in,II 、c out,II and c out,II is the specific heat capacity of the air at the inlet and outlet of the two fans; u air,I and u air,II It is the flow rate of primary air and secondary air.

[0202] Heat analysis of the outflow from the incinerator:

[0203]

[0204] Among them, among them, Q steam Indicates the heat of steam generation, Q dis,heat Indicates the heat loss of the furnace body, Q inc,solid Indicates the heat loss from incomplete combustion of solid phase, Q inc,gas Indicates the heat loss from incomplete combustion in the gas phase, Q slag,loss represents the sensible heat loss of the slag,Q gas,loss Represents flue gas heat loss. Its calculation formula is as follows:

[0205]

[0206] in,, h sh and h ws are the superheated steam enthalpy and feed water enthalpy. S rated and S is the rated main steam flow and actual steam flow. Q inc,carbon and Q inc,volatile is the heat of unburned fixed carbon and unvaporized volatile matter, m unre,CO and m unre,CH4 is the mass of unreacted gas, A ar is the proportion of ash in the fuel, c slag is the specific heat capacity of the slag, φ slag,carbon is the carbon content of the slag, T slag and T gas are the slag temperature and flue gas temperature, T 0 is the ambient temperature, m En is the mass of the last block in E, is the average time it takes for each grate piece in E to move to the next position, c gas is the average specific heat capacity of the flue gas.

[0207] Energy balance model used to calculate furnace temperature T furnace , main steam flow F steam and flue gas oxygen content O 2%, where for example the furnace temperature can be estimated by the following relationship:

[0208]

[0209]

[0210] in, T in Indicates the temperature of the air at the furnace inlet, Q loss represents the sum of the total heat losses, mair Indicates the total mass flow rate of air.

[0211] As an example, the static model is converted into a dynamic model, and time factors and real-time variables are introduced into the model so that the model can reflect the instantaneous changes in the incineration process. For example, the furnace temperature change can be described as:

[0212]

[0213]

[0214] in, Q net (t) is the time t Net heat input; Q in (t) and Q out (t) are the energy entering and leaving the incinerator at time t; Q loss (t) is the total loss of the system at time t; m is the total mass of substances involved in heat transfer in the system; c p is the specific heat capacity.

[0215] Online parameter calibration: Setting key parameters in dynamic models (such as heat transfer coefficient, reaction rate constant, etc.), and perform periodic calibration through real-time data collection. Let the calibration function be:

[0216]

[0217] in: θ exp ( t ) is the parameter obtained by fitting the real-time monitoring data; λ is the calibration step factor; Δ t is the calibration time interval.

[0218] The model parameters are periodically calibrated using real-time collected data to ensure that the model is highly consistent with the actual process.

[0219] Model verification and simulation experiments verify the dynamic modeling results through experimental data to ensure that the model's simulation accuracy of the physical and chemical processes in each area meets the requirements.

[0220] Offline verification:

[0221] Compare the model's predicted values ​​for key indicators such as air volume, temperature, and steam flow with the actual measured values. Error evaluation metrics can be defined, such as the root mean square error (RMSE):

[0222]

[0223] in, X model,i and X exp,i They represent the dynamic model and any key output variable of real-time monitoring (such as furnace temperature, air volume, etc.).

[0224] Simulation experiment

[0225] The dynamic model is simulated using experimental data to verify the accuracy of the model's simulation of material transfer, heat exchange, and chemical reaction processes within each working area. The absolute or relative error of each indicator prediction is ensured to be within the preset tolerances.

[0226] Based on the validation results, adjust the model parameters and constants in the energy balance formula for each region until the ideal prediction is achieved. This solidifies the dynamic modeling solution for subsequent integrated control. Compare the dynamic model's predicted air volume, temperature, steam flow, and other indicators with the actual measured values ​​to ensure that the error is within the preset range.

[0227] Figure 7 、 Figure 8 、 Figure 9 The figures show a comparison between the actual measured values ​​and the model output for furnace temperature, main steam flow rate, and flue gas oxygen content. The errors in the main steam flow rate and furnace temperature compared to the actual measured values ​​are both less than 3%. Flue gas oxygen content fluctuates significantly, with a model-calculated error of 8.02%. This demonstrates the model's high reliability in accurately predicting key operating parameters. Even with large fluctuations in flue gas oxygen content, the model maintains reasonable prediction accuracy under complex operating conditions. Therefore, the integrated model control system optimizes input parameters to ensure that the key outputs of the MSWI meet operational requirements.

[0228] Step 4: Build an integrated control system, fuse the predicted MSWI critical air volume with the simulated theoretical air volume, obtain a comprehensive prediction value through a weighted method, and use the comprehensive prediction value to perform integrated control of the MSWI air volume. This includes: In the integrated control system, the machine learning model adjusts the predicted air volume based on real-time operation data, while the mathematical and physical model adjusts the grate movement speed based on the demand analysis of the air volume calculated by chemical reactions and thermodynamics. v , air volume proportional coefficient cr and excess air coefficient n , to adapt to actual changes.

[0229] Specifically, if Figure 4 As shown in the figure, the results obtained by Bio-HybridNet (including the spiking neuron module, LSTM module and synaptic plasticity module) and the incineration process dynamic modeling module are fused, including:

[0230] Build a data fusion interface and obtain comprehensive prediction values ​​through weighted methods V fused :

[0231]

[0232] in, λ ∈[0,1] is the fusion weight, which is used to balance the contribution of machine learning prediction results and physical model calculation results.

[0233] As an embodiment, in order to achieve accurate adjustment of the air volume control system, the objective function is set J ( v , n , cr ), the objective function can be defined as the mean square error between the predicted value and the theoretical value:

[0234]

[0235] in, v Indicates the grate movement speed; n Indicates excess air coefficient; cr Indicates the air volume proportional coefficient;

[0236] V pred is the predicted actual wind volume; V theory The theoretical air volume calculated by the physical model.

[0237] At the same time, the primary air volume, secondary air volume, grate speed, excess air coefficient and air volume proportional coefficient are taken as key indicators of system control, and their sensitivity to the error is analyzed.

[0238] As an example, the dynamic parameter adjustment mechanism dynamically adjusts the control parameters through real-time optimization methods based on the results of offline and online tuning. Taking the gradient descent method as an example, each parameter is updated separately:

[0239]

[0240] in, η v 、 η n and η cr are the online learning rate or update step size of the corresponding parameters respectively; 、 and is the partial derivative of the objective function with respect to each parameter, reflecting the impact of each parameter on the prediction error. After the update, the new parameter values ​​are immediately put into the actual control system, and closed-loop adaptive adjustment is achieved by solving the minimization of the objective function J.

[0241] Real-time feedback and closed-loop control: Establish a real-time monitoring system to collect key data such as furnace temperature, steam flow, flue gas oxygen content, and pollutant emissions, and record the actual measured air volume V measured .

[0242] Using feedback error E ( t ) is used to calibrate the control model, and the error is defined as:

[0243]

[0244] The feedback error is used as the basis for dynamic adjustment to further adjust the control parameters and form a closed-loop feedback control. E ( t )> E th (preset threshold), the system starts the parameter adjustment process again to ensure that the error is reduced and kept within the preset range through real-time adjustment.

[0245] System Verification and Effectiveness Evaluation: The integrated control model is tested over a long period of time to evaluate the system's response speed and robustness under different operating conditions, as well as the degree of agreement between the predicted and actual air volume. The model's prediction accuracy is evaluated using goodness of fit (R²), root mean square error (RMSE), mean absolute percentage error (MAPE), and maximum absolute error (MAXE), using the following formula:

[0246]

[0247]

[0248]

[0249]

[0250] in, y i is the true value, is the predicted value, is the average of the true values, is the average of the predicted values, N is the sample size.

[0251] According to the requirements of the MSWI control process, the furnace temperature, main steam flow rate, and flue gas oxygen content are used as the output of the model. The efficiency is calculated using the following formula:

[0252]

[0253]

[0254]

[0255] Where, m air is the total amount of air, m surplus,O2 is the residual oxygen content after the reaction, u induce is the flow rate of the induced airflow. q MSW is the lower calorific value of MSW, kJ / kg.

[0256] Further analysis of energy efficiency, environmental protection effects and the stability of key indicators (such as grate speed, excess air coefficient, and air volume proportional coefficient) will ensure that the overall system performance meets the expected goals, and ultimately form a complete MSWI air volume intelligent control solution to achieve integrated control of MSWI components.

[0257] Figure 10 and Figure 11 The figure shows a comparison between the predicted primary and secondary air volumes and the actual air volume values ​​for the integrated control model. As can be seen from the figure, after optimization, the model has a high degree of fit. The predicted air volume values ​​maintain a high degree of consistency with the theoretical calculated values ​​in terms of overall trends, with an error within 5%. This demonstrates that the model can accurately capture the changing patterns of air volume. In areas of large fluctuations, the predicted values ​​can dynamically adjust to the theoretical calculated values, demonstrating that the model has good responsiveness to fluctuations in air volume. A dynamic adjustment relationship exists between the primary and secondary air volumes; as the primary air volume increases, the secondary air volume decreases, meeting actual operational requirements.

[0258] The fluctuation range of flue gas oxygen content is 4.12%~9.45%. The oxygen content within this range indicates that the garbage has received a relatively sufficient oxygen supply during the incineration process, the combustion efficiency is high, and the generation of incomplete combustion products is also reduced. Compared with the actual detection values, the errors of the main steam flow and furnace temperature output by the model are less than 3%. The flue gas oxygen content fluctuates greatly, and the error of the model calculation is 8.02%. This reflects that the model has high reliability in the accurate prediction of key operating parameters. Even in the face of large fluctuations in flue gas oxygen content, the model can still maintain reasonable prediction accuracy under complex working conditions, reflecting its adaptability and robustness. The emission concentration of CO is stably lower than , The emission concentration is stable below , which is far below the standard limit ( , This shows that the model can effectively optimize the incineration process, significantly improve the degree of combustion completeness, reduce CO emissions, and reduce the amount of CO emissions by precisely controlling the combustion temperature and air volume distribution. The generation of this model demonstrates the significant advantages of dynamic models integrated with machine learning models in environmental performance and combustion process optimization.

[0259] Example 2

[0260] In one embodiment of the present disclosure, a MSWI air volume control system based on a bionic hybrid model and dynamic modeling is provided, comprising:

[0261] The signal acquisition module is used to obtain the continuous key working condition signals of the MSWI system and convert them into pulse sequences after preprocessing;

[0262] The feature extraction and air volume prediction module is used to input the pulse sequence into the optimized biomimetic hybrid neural network model. The pulse signal vector of the pulse sequence is first extracted through the pulse neural network module, and then the pulse signal vector is transmitted to the LSTM module. During the transmission process, a short-term plasticity mechanism is introduced. By introducing a dynamic weight attenuation coefficient and recovery coefficient, the response of synaptic transmission is adjusted to the sudden input. A long-term plasticity mechanism is introduced between the input connection and the gate control unit within the LSTM module. The network weight is persistently adjusted for positive and negative feedback information. Finally, the LSTM module predicts and outputs the MSWI critical air volume at the future moment.

[0263] The dynamic simulation module is used to divide the MSWI incineration process into multiple zones, establish corresponding material balance, energy balance and reaction rate mathematical models for each zone, simulate the physical and chemical reaction process of waste incineration, and obtain the simulated theoretical air volume;

[0264] The integrated control module is used to fuse the predicted MSWI key air volume with the simulated theoretical air volume, obtain the comprehensive prediction value through the weighted method, and use the comprehensive prediction value to perform MSWI air volume integrated control.

[0265] Example 3

[0266] In one embodiment of the present disclosure, a computer program product is provided, including a computer program. When the computer program is executed by a processor, the MSWI air volume control method based on the bionic hybrid model and dynamic modeling is implemented.

[0267] Example 4

[0268] In one embodiment of the present disclosure, a non-transitory computer-readable storage medium is provided, which is used to store computer instructions. When the computer instructions are executed by a processor, the MSWI air volume control method based on the bionic hybrid model and dynamic modeling is implemented.

[0269] Example 5

[0270] In one embodiment of the present disclosure, an electronic device is provided, including: a processor, a memory, and a computer program; wherein the processor is connected to the memory, and the computer program is stored in the memory. When the electronic device is running, the processor executes the computer program stored in the memory, so that the electronic device executes the MSWI air volume control method based on the bionic hybrid model and dynamic modeling.

[0271] The present disclosure is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present disclosure. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0272] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0273] Although the above describes the specific implementation methods of the present disclosure in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present disclosure. Those skilled in the art should understand that on the basis of the technical solution of the present disclosure, various modifications or variations that can be made by those skilled in the art without creative work are still within the scope of protection of the present disclosure.

Claims

1. The MSWI air volume control method based on bionic hybrid model and dynamic modeling is characterized by: include: Obtain the continuous key operating condition signals of the MSWI system, and convert them into pulse sequences after preprocessing; The pulse sequence is input into the optimized biomimetic hybrid neural network model. The pulse signal vector of the pulse sequence is first extracted through the pulse neural network module. The pulse signal vector is then transferred to the LSTM module. During the transfer process, a short-term plasticity mechanism is introduced. By introducing a dynamic weight attenuation coefficient and recovery coefficient, the response of synaptic transmission is adjusted to the sudden input. A long-term plasticity mechanism is also introduced between the input connection and the gate control unit within the LSTM module. The network weight is persistently adjusted for positive and negative feedback information respectively. Finally, the LSTM module predicts and outputs the MSWI critical air volume at the future moment. The MSWI incineration process is divided into multiple zones, and corresponding material balance, energy balance and reaction rate mathematical models are established for each zone to simulate the physical and chemical reaction processes of waste incineration and obtain the simulated theoretical air volume. The predicted MSWI critical air volume is fused with the simulated theoretical air volume, and a comprehensive prediction value is obtained through a weighted method. The comprehensive prediction value is used for integrated control of MSWI air volume.

2. The MSWI air volume control method based on bionic hybrid model and dynamic modeling according to claim 1 is characterized in that: The key operating condition signals include air volume, temperature, furnace temperature, air inlet pressure, flue gas CO concentration and flue gas NO X Concentration, the preprocessing process includes denoising and normalization processing, and the corresponding pulse frequency is set for the key working condition signal after preprocessing according to the signal amplitude. When the signal value is large, a high-frequency pulse is generated, and when the signal value is small, a low-frequency pulse is generated, thereby mapping the continuous working condition signal into a discrete pulse sequence mode.

3. The MSWI air volume control method based on bionic hybrid model and dynamic modeling according to claim 1 is characterized in that: The bionic hybrid neural network model includes a pulse neural network module, an LSTM module and a synaptic plasticity module. The pulse neural network module includes pulse neurons and multi-layer SNN convolution. The pulse neurons are designed based on the LIF model, and the LIF model is used to describe the membrane potential changes of the pulse neurons; the multi-layer SNN convolution network adopts a convolution kernel and a time delay window with bionic settings in each layer to simulate the response of the biological visual cortex to signals of different frequencies. A time domain pooling layer is set after each convolution layer. By performing local statistics on the pulse sequence and extracting local time features, the output pulse sequence obtained after multi-layer pulse convolution and pooling processing is converted into a feature vector of fixed dimension to obtain a pulse signal vector.

4. The MSWI air volume control method based on bionic hybrid model and dynamic modeling according to claim 1 is characterized in that: The LSTM module includes a multi-layer LSTM network, each layer contains several LSTM neurons. The first layer is used to capture preliminary temporal dependencies, and subsequent layers are used to extract deep time series. In order to simulate the learning ability of the biological brain, a synaptic plasticity module is introduced. The synaptic plasticity module includes a short-term plasticity mechanism and a long-term plasticity mechanism. The short-term plasticity mechanism is introduced between the multi-layer SNN convolution and the LSTM module, and the long-term plasticity mechanism is introduced inside the LSTM module. The short-term plasticity mechanism simulates the rapid consumption and recovery of neurotransmitters between synapses, introduces a short-term weight adjustment factor, and modulates signal transmission, so that the network shows higher sensitivity to sudden and instantaneous input changes; the long-term plasticity mechanism adopts a weight correction algorithm based on error feedback, so that when the network prediction error is large, the relevant synaptic weights are corrected multiple times to achieve long-term enhancement or inhibition, so as to improve the memory effect of the overall model.

5. The MSWI air volume control method based on bionic hybrid model and dynamic modeling according to claim 1 is characterized in that: The MSWI incineration process is divided into multiple areas, including the hopper area, feeder area, drying area, combustion area, solid-phase combustion area, ash area and waste heat boiler area. The material transportation, heat exchange and chemical reaction processes in each area are modeled, and corresponding material balance, energy balance and reaction rate mathematical models are established respectively. In addition, calculation formulas for material flow and chemical reaction in each area are established, including the distribution calculation of primary and secondary air in each area, thermal radiation heat transfer, convection heat transfer and chemical reaction conditions. Time factors and real-time variables are introduced into the model to enable the model to reflect the instantaneous changes in the incineration process and obtain the simulated theoretical air volume.

6. The MSWI air volume control method based on bionic hybrid model and dynamic modeling according to claim 1, characterized in that: A data fusion interface is constructed, and the predicted MSWI key air volume and the simulated theoretical air volume are fused through the weighted method. An integrated control model is established, and the fused data is input into the integrated control model. The primary air volume, secondary air volume, grate speed, excess air coefficient and air volume proportional coefficient are used as key control indicators, and closed-loop adaptive adjustment is achieved by solving the minimization objective function.

7. The MSWI air volume control system based on bionic hybrid model and dynamic modeling is characterized by: include: The signal acquisition module is used to obtain the continuous key working condition signals of the MSWI system and convert them into pulse sequences after preprocessing; The feature extraction and air volume prediction module is used to input the pulse sequence into the optimized biomimetic hybrid neural network model. The pulse signal vector of the pulse sequence is first extracted through the pulse neural network module, and then the pulse signal vector is transmitted to the LSTM module. During the transmission process, a short-term plasticity mechanism is introduced. By introducing a dynamic weight attenuation coefficient and recovery coefficient, the response of synaptic transmission is adjusted to the sudden input. A long-term plasticity mechanism is introduced between the input connection and the gate control unit within the LSTM module. The network weight is persistently adjusted for positive and negative feedback information. Finally, the LSTM module predicts and outputs the MSWI critical air volume at the future moment. The dynamic simulation module is used to divide the MSWI incineration process into multiple zones, establish corresponding material balance, energy balance and reaction rate mathematical models for each zone, simulate the physical and chemical reaction process of waste incineration, and obtain the simulated theoretical air volume; The integrated control module is used to fuse the predicted MSWI key air volume with the simulated theoretical air volume, obtain the comprehensive prediction value through the weighted method, and use the comprehensive prediction value to perform MSWI air volume integrated control.

8. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the MSWI air volume control method based on the bionic hybrid model and dynamic modeling according to any one of claims 1 to 6 is implemented.

9. A non-transitory computer-readable storage medium, characterized in that The non-transitory computer-readable storage medium is used to store computer instructions. When the computer instructions are executed by the processor, the MSWI air volume control method based on the bionic hybrid model and dynamic modeling as described in any one of claims 1 to 6 is implemented.

10. An electronic device, characterized in that: include: A processor, a memory, and a computer program; wherein the processor is connected to the memory, and the computer program is stored in the memory. When the electronic device is running, the processor executes the computer program stored in the memory to enable the electronic device to implement the MSWI air volume control method based on the bionic hybrid model and dynamic modeling as described in any one of claims 1 to 6.

Images