Frontiers in Computing and Intelligent Systems 
ISSN: 2832-6024 | Vol. 3, No. 1, 2023 

 

117 

Power Supply Coal Consumption Prediction Model 
Based on Long and Short-term Memory Neural Network 

Mengjie Liu*, Chenggang Zhen 

School of Control and Computer Engineering, North China Electric Power University, Baoding 071003, China 

* Corresponding author: Mengjie Liu 

 

Abstract: As a measure of the overall economic performance of thermal power plants and an important component of the 

variable cost of power generation, the prediction of coal consumption of power supply is of great significance for the bidding 

decision of power generation companies. In this paper, we propose an improved long and short-term memory (LSTM) model for 

calculating the coal consumption of thermal power units, using a large amount of data stored in the plant monitoring information 

system (SIS). Firstly, the data are pre-processed using the threshold determination method to filter out the data of stable operating 

conditions. Secondly, the maximum correlation minimum redundancy (mRMR) algorithm is used to determine the optimal set 

of special features. The results show that the prediction effect of the improved LSTM-based power supply coal consumption 

calculation model proposed in this paper is better than other models, and the calculation accuracy is higher, which is suitable for 

power supply coal consumption calculation. 

Keywords: Power supply coal consumption; MRMR; LSTM. 

 

1. Introduction 

At present, the real-time calculation of power supply coal 

consumption mainly adopts data mining method. Based on 

the support vector machine technology, the literature [1] 

analyzed the sensitivity coefficients of the operating 

parameters on the coal consumption under different load 

conditions and provided guidance for the optimal operation 

adjustment of the unit. In the literature [2], the BP algorithm 

was used to establish the prediction model of coal 

consumption with strong dynamic response performance, but 

the coal consumption characteristics under different load 

conditions were not further studied. In the literature [3], the 

energy efficiency indexes that have a major impact on the coal 

consumption of power supply were selected by using the gray 

correlation analysis method and combined with the 

generalized regression neural network to predict the coal 

consumption of power supply. In the literature [4], a discrete 

model between load and coal consumption rate was 

established by using fuzzy clustering, rough set theory and 

other data mining methods to predict the current coal 

consumption characteristics of the unit and achieve the 

optimal load allocation of the unit in real time. In the literature 

[5], the mRMR algorithm was used to filter the optimal subset 

of features by ranking the importance of the initial input 

variables, which improved the prediction accuracy of the 

model. In the literature [6], a black box model was established 

based on the long short-term memory (LSTM) neural network, 

and the comparative validation showed that the LSTM model 

outperformed the SVR and BP models in all evaluation 

indexes. 

In this paper, a 600MW domestic supercritical coal-fired 

unit is used as the research object, the stable working 

condition data is screened by the threshold determination 

method, the optimal set of special features is determined by 

the mRMR algorithm, and finally the power supply coal 

consumption prediction model is established based on LSTM. 

The application of the algorithm model was carried out with 

an ultra-supercritical million wet-cooled unit as the research 

object. 

2. Methodology 

2.1. Screening of steady state data 

When the unit is in variable operating conditions, 

fluctuations in operating parameters occur. Therefore, the 

collected data need to be pre-processed, i.e., stability 

judgment, and the data of non-stable operating conditions 

need to be excluded[7]. The determination of the stable 

operating conditions of the unit needs to be based on the 

important operating parameters that characterize the 

operating status of the unit. If the unit load, feedwater flow, 

reheat steam temperature, exhaust smoke temperature, and 

other parameters are all tested for steady state, sufficient 

stable data will not be available for model training. Therefore, 

in this paper, unit load, main steam pressure and reheat steam 

temperature are selected for steady-state condition 

determination. 

The data for the steady-state condition were screened by 

the threshold method. Within 60 min, the difference between 

the maximum and minimum values of the unit load, main 

steam pressure and reheat steam temperature is less than a 

certain threshold, which is considered to be in steady-state 

conditions. Otherwise, the sampling slide window is pushed 

back clockwise for 5 min, and the steady state is re-judged 

until the sample meets the requirements. 

2.2. MRMR-based key feature extraction 

In probability theory and information theory, the concept 

of mutual information is introduced to measure the degree of 

interdependence between two random variables, reflecting 

both linear and nonlinear correlations between the two sets of 

variables. For 2 sets of discrete random variables X and Y , 

the mutual information between them is ( ; )I X Y . 

The formula is as follows : 

2

( , )
( ; ) ( , )log

( ) ( )x X y Y

p x y
I X Y p x y

p x p y 

=       (1) 



 

118 

The mRMR calculates the correlation between each input 

variable and the output variable based on the mutual 

information theory, while taking into account the correlation 

between each input variable. The mRMR algorithm can find 

the subset of features from the initial input variables that have 

the highest correlation with the output variables and the 

lowest correlation between the input variables. 

( ) ( )2
,

1 1
max ; ;

i n i j n

i i j

X S X X S

I X Z I X X
n n 

 
− 

 
 
          (2) 

An incremental search is used to sequentially select the 

optimal feature subset, which satisfies the following equation. 

( ) ( )
1

1
max ; ;

1
j m

i i j

X S

I X Z I X X
m

−

 
− 

 − 
∣          (3) 

The mRMR algorithm was used to rank the degree of 

influence of all features and added to the feature subset in 

order. To eliminate the difference in magnitude between 

variables, the sample data need to be normalized by the 

formula: 

min

max min

i
i

x x
M

x x

−
=

−
                (4) 

The final feature subset is best predicted by the prediction 

model when unit load, exhaust oxygen, condenser vacuum, 

total coal feed, ambient temperature, and low-level heat of 

incoming coal are taken. 

2.3. LSTM-based coal consumption prediction 

for power supply 

Long-term memory in learning is difficult to be preserved 

due to the gradient explosion or gradient disappearance 

problem of Recurrent Neural Network (RNN).In 1997, 

Hochreiter et al. first proposed Long-Term and Short-Term 

Memory Network, LSTM)[8]. 

LSTM is developed on the basis of RNN recurrent neural 

network, which is a gated recurrent neural network.The 

internal structure of LSTM is shown in Figure 1.The key of 

LSTM is the memory cells, which have states throughout the 

network with only a small amount of information interaction, 

making the information stable during the long-term 

propagation.LSTM is implemented through input gates, 

forgetting gates and output gates to achieve protection and 

control of gradient information. Where the forgetting gate 

determines which information is discarded or retained from 

the cell state; the input gate updates the cell state; and the 

output gate determines the final output result. 

 

Figure 1. The structure of LSTM network 

The equation for calculating each part of the neuron at time 

t is shown below. 

( )1t xf t hf t ff W X W h b −= + +                   (5) 

( )1t xi t hi t ii W X W h b −= + +                     (6) 

( )1

1

tanh

* *

t xc t hc t c

t t t t t

c W X W h b

c f c i c



−



−

 = + +


= +
                (7) 

( )1t xo t ho t oo W X W h b −= + +                (8) 

( )*tanht t th o c=                         (9) 

3. Results and discussion 

Figure 2 shows the comparison of the prediction effect of 

LSTM with error back propagation (BP) and support vector 

machines (SVM). The structure of LSTM in this paper is: 3-

layer structure, optimizer is Adam, loss function is MSE, 

number of nodes in hidden layer is 20, learning rate is 0.001. 

100 iterations. Compared with BP and SVM, LSTM has better 

prediction results.  

 

Figure 2. Prediction effect of three models 

Table 1 shows the comparison of the prediction accuracy 

of the two networks. 

Table 1. Comparison of prediction accuracy 

models MAE RMSE MAPE% 

LSTM 7.619 10.552 2.321 

BP 8.274 11.520 2.539 

SVM 8.505 11.919 2.610 

The calculated value of LSTM model of coal supply 

consumption of the unit is in good agreement with the actual 

value. The low load operation of the unit is unstable and the 

parameters fluctuate greatly, resulting in a relatively high 

error in the calculation of power supply coal consumption. 

The experimental results show that the prediction model used 

in this paper can achieve the calculation of power supply coal 

consumption in thermal power plants better.  

4. Conclusion 

LSTM is an efficient deep learning model that can handle 

large amounts of data efficiently. It is proven to be more 

accurate and applicable by comparing with other prediction 

methods. For the selection of the input parameters of LSTM, 

a maximum small-off minimum redundancy method is used 

to select the influences with high relevance to form the 

optimal set of special features. The prediction error is less 

than 3%, and the accuracy of the predicted coal consumption 

of thermal power units is high. 

References 

[1] Liu B.H., Fu Z.G., Wang P.K., et al (2018). Research on the 
application of big data mining data technology in the analysis 
of energy consumption of coal-fired power plant units. 
Proceedings of the CSEE, vol.12, no.38, p.3578-3587. 



 

119 

[2] Du G., Liao L, Zhang D.L., Yang K (2011). Online simulation 
of coal consumption in thermal power plants based on BP 
neural network. Computer Simulation, vol.4, no.28, p.313-344. 

[3] Xu J.,Gu Y.J.,Wang Z.,Li Q.Q.,Yang N (2017). Research on 
energy efficiency characteristic indexes of coal power units and 
their benchmark values based on data mining. Proceedings of 
the CSEE, vol.7, no.37, p. 2009-2016. 

[4] Zeng D.L., Yang T.T., Cheng X, et al (2010). Application of 
data mining methods in real-time plant-level load optimization 
allocation. Proceedings of the CSEE, vol.11, no.30, p. 109-114. 

[5] Jin X.Z.,Liu Y.,Zhao W.J.,Yu J (2022). Prediction of SO_2 
mass concentration at the outlet of thermal power plants based 
on mRMR and MA-RELM. Journal of Chinese Society of 
Power Engineering, vol.7, no.42, p.664-676. 

[6] Zheng T.X., He J., Zhang L.B (2020). LSTM neural network-
based combustion timing estimation for hybrid fuel HCCI 
engines under complex operating conditions. Chinese Journal 
of Scientific Instrument, vol.10, no.41, p.100-110. 

[7] Opri I, Victor-Eduard C , Norior M , et al (2020). Parametric 
optimization of the thermodynamic cycle design for 
supercritical steam power plants. Energy Conversion and 
Management, no.208, p. 112587. 

[8] Liu Y, Li Q, Duan X, Zhang Y, Yang Z, Che D (2018). 
Thermodynamic analysis of a modified system for a 1000 MW 
single reheat ultra-supercritical thermal power plant. Energy, 
no.145, p. 25–37.r