Performance evaluation and artificial neural network based thermal modelling of multi tilted active tiles in data centres

Funding information This work was funded in part by the National Natural Science Foundation of China (NSFC) under Grants. 61862048, 61762070, and 61962045, Inner Mongolia Key Technological Development Program (2019ZD015), and Key Scientific and Technological Research Program of Inner Mongolia Autonomous Region (2019GG273, 2020GG0094, NJZY21321). Abstract The main purpose of this paper is to improve the resilience of data center cooling system and reduce the energy consumption in data center by using the Tilted Active Tile (TATs). We deploy multiple TATs in cold aisle environment then analyze the thermal performance of racks. TATs also help to improve the reliability in case of cooling failure by increased ride through time. We observed an extended ride-through time for data centers deployed with Tilted Active Tiles (TATs) when the CRAC blower failed. Proposed approach can maintain cooling system temperature for up to 10 minutes. We also introduced ANN based thermal models to predict the thermal performance of TATs, To train the ANN models, we collect the rack inlet and CRAC outlet temperature from Data Center. We also fixed blank panel in empty U part of racks and then analyzed that the thermal performance of racks are better and temperature distribution at bottom, middle and top of the rack were same under 30angle. The airflow of TATs and passive tile influence on surrounding racks. That’s why we analyzed the disturbance of TATs on surrounding rack and observed that the TATs at 30angle has positiveimpact on the surrounding racks.


INTRODUCTION
Thermal management system of data centre is one of the major problems that has emerged in recent years, which is employed extensively for data storage and retrieval. It has brought an efficient and convenient lifestyle for people, but the resilience of data centre and increasing energy consumption of data centre are huge problems.
In this paper we proposed TATs to improve the resilience of data centre problem and reduced the energy consumption from data centre. Data centre is a large scale computing system which requires extremely high level of reliability, while the proposed TAT improves the cooling efficiency and also enhances the reliability by increasing the ride through time in case of cooling failure. For example in case of cooling failure in data centre TAT can provide cooling up to 10 minutes and passive tile stop cooling on the spot. The reason for this is that there is available thermal mass in the room and/or air circulation and tiles fans can partially drive the airflow, thus improve the thermal condition even if the CRAC blower malfunctions [21]. To reduce the energy consumption from data centre it is necessary to study cooling system in data centre. The cooling system consumed 30-55% energy in a data centre, the cooling infrastructure mainly used for removing the heat dissipation from servers in data centre to ensure server performance and reliability [21]. In raised-floor data centres, the cold air flows through the perforated tile into the cold aisle, and enters server racks to absorb waste heat generated by servers, and finally be ejected into the hot aisle [21]. Perforated tiles generally can be classified as passive tiles and active tiles. Passive tiles are the ordinary tile in data centre where the cold air flow is driven by the pressure differential between over-and underfloor spaces. Active tiles have attached fans which can actively draw the cold air from the underfloor plenum.
In previous papers, researchers focused on traditional active tiles, where tile fans are placed horizontal to the floor, as shown in (Figure 1(b)). Because of tiled fans placed horizontally sometimes the airflow blows over the rack and does not enter into the rack. To solve this problem, we propose Tilted Active Tiles (TATs), in which the tile fans are tilted so that the cold air flow can be directly blow into the racks, rather than blow over the rack. In TATs we change the angle of the tile from 0 • to 40 • , every 5 • a span. We adjusted multi tilted active tile at different position in cold aisle environment and analysed the thermal performance of the racks. We analysed the thermal performance of the proposed TATs via experiment. In particular, we adjusted active tiles at different layout and angle as shown in (Figure 1(a)), then compared the thermal performance under different environment. The main purpose of this experiment is to analyse the thermal performance of TATs. Firstly, the upper, middle and lower positions of the racks are analysed. We found that the most of the temperature changed at middle and bottom of the rack when we adjust the TAT at 30 • angles. We also proposed the black panels to fill the gap at the bottom of the rack and the empty U part of the rack and then analysed the performance of racks. Because empty part of the rack also cause the hot air circulation. We analysed that the performance of the TATs best at 30 • angle. But the gap at the bottom of the rack and the empty U part of the rack cause the hot air circulation. So that is why we close the gap at the bottom of the rack with black panels, and repeated the experiment. We also analysed that when we adjust TATs at 30 • angle it has positive impact on the surrounding racks, but with increasing distance of rack from the tilted active tile, weaken the impact, the closer the distance between the tilted active tile and rack and increase the fan speed has negative impact on racks.
In this paper we further introduced the ANN model. Thermal efficiency model was established by BP and LSTM, in this experiment single output model and multi output model were analysed. The test set error of the single output structure is greater than the training set and validation set. The predicted effect of BP model is better than LSTM. The average prediction error is 0.57. In multi output model LSTM prediction error is less than BP. The average prediction error is 0.07.

RELATED WORK
There have been a number of existing surveys on thermal related issues for data centers. Here we group these works into three categories. Thermal modeling, performance evalua-tion, and reduced the energy consumption. The thermal performance can be evaluation, and reduced the energy consumption. The thermal performance can be evaluated by set of performance of tilted active tile (TATs). Some work in this area include [2]. Which summarized the thermal performance of rack through active tile. Athyale and Joshi [2] presented some thermal measurement tools which are commercially available or developed by themselves. They also analyzed the performance of perforated tiles and containment systems. Jiangxiong wan [21] and Pramod Kumar [11] reported research results for ventilation systems, underfloor plenum, perforated tiles, and rack layouts, etc. A set of design guidelines were also recommended. They also discussed the factors affecting the air flow distribution in underfloor plenum, server rack, and machine room. Patterson [14] reviewed techniques and codes/regulations for the containment system. Although these surveys provided detailed reviews on various kinds of air flow management techniques, none of them covered this topic from the perspective of whole air flow cycle. Our work is intended to bridge this gap and to incorporate the most recent works in this field. Ni and Bai [13] also reviewed thermal performance and presented evaluation standards. The main target of thermal modeling is to build the two different model and then compare the performance of these two model. Some work in this area include [20]. Which summarize and organized a large body of energy model for both cooling and IT system There have been a number of existing surveys on minimize the energy consumption in data center. To address this problem various thermal aware scheduling were proposed, which were reviewed by Chaudhry et al [3] and Kong and Liu [9]. Thermal, electrical and energy management challenges are discussed in [6] [10]. The second perspective is the cooling system solution seeking to improve the cooling efficiency for a given heat load. Literatures [13], [1] [18] provided reviews for thermal management techniques in traditional air-cooled data centers. Other advanced topics in this area include liquid and hybrid cooling [1], [8], [5], [7], [15] free cooling [21], and waste heat recovery [4], [19], [12], etc. There are also some surveys which covered both perspectives, such as [17] and [16]- [6]. While most of works discussed the two perspectives independently, an apparent fact is that IT and cooling systems are coupled via heat. Treating them separately may only lead to limited impacts [13] and suboptimal solutions [1] [21]. As a consequence, joint optimization techniques which were reviewed in [21] become a cutting-edge research direction.

Experiment setup
All experiments are conducted in a data centre of Inner Mongolia Meteorology Information Centre.  There have been a number of existing surveys on thermal related issues for data centres. Here we group these works into three categories. Thermal modelling, performance evaluation, and reduced the energy consumption. The thermal performance can be evaluated by set of performance of tilted active tile (TATs). Some work in this area include [2]. Which summarized the thermal performance of rack through active tile. Athyale and Joshi [2] presented some thermal measurement tools which are commercially available or developed by themselves. They also analysed the performance of perforated tiles and containment systems. Jiangxiong Wan [21] reported research results for ventilation systems, underfloor plenum, perforated tiles, and rack layouts etc. A set of design guidelines were also recommended. They also discussed the factors affecting the air flow distribution in underfloor plenum, server rack, and machine room. Patterson [14] reviewed techniques and codes/regulations for the containment system. Although these surveys provided detailed reviews on various kinds of air flow management techniques, none of them covered this topic from the perspective of whole air flow cycle. Our work is intended to bridge this gap and to incorporate the most recent works in this field. Ni and Bai [13] also reviewed thermal performance and presented evaluation standards. The main target of thermal modelling is to build the two different model and then compare the performance of these two model. Some work in this area include [20], which summarize and organized a large body of energy model for both cooling and IT system There have been a number of existing surveys on minimize the energy consumption in data centre to address this.

Thermal performance measurement
We build a temperature measurement system using some measurement tools to measure and collect the inlet temperature of the rack and outlet temperature of the CRAH. The measurement system consists of six sensors in different heights on the rack. We collected the thermal data for TATs and ordinary AVTs to compare their performance. Firstly, collect the inlet temperature of the rack under the condition of tilted active tile. In this experiment we change the angle of the tilted active tile. Tilted active tile angle range is 0 • to 40 • , 5 • every one span. In this experiment, the PWM range is 0 • to 250 • , and every 10 PWM is one step, every change the PWM need to stay for two minutes, makes the current environment tends to be stable. When cold aisle environment reach static state and 200 pieces of steady state data were collected, then we changed the PWM, The time of each data collection is approximately 2.4 s. We used hand-held anemometer to collect the outlet temperature of the CRAH. We adjust hand-held anemometer vertically below the passive tile. Where the cold aisle is closest to the CRAC. The wind speed measurement range is 0 45 m/s, the temperature range is 0-45 • C and the accuracy is ±0.3 and ±0.1 • C respectively.

Performance evaluation of TATs at different angles
The main purpose of this experiment is to analyse the thermal performance of TATs. We adjusted TATs at different angle, and compared the thermal performance of TATs under different environment. We adjust the TATs angle from 5 • to 40 • range. Every 5 • is a span. Firstly analysed the temperature at the upper, middle and bottom position of the racks and then we analysed the temperature distribution of the racks under different angles of TATs. First we compared the rack temperature under different condition of the TATs. As we can see the ( Figure 3) if the rack temperature is greater than 0 then it means the temperature of the rack is decreasing and if the temperature is lower than 0, it means temperature of the rack is increasing, which is caused of hot spot. When the tilted active tile angle is 0, and increases PWM from 50 to 100, the temperature of the bottom and middle points on the rack changed from 0.368 to −0.317°C, which means temperature of the rack increases, which is caused of hot spot. As we can see in Figure 3 that the temperature of the rack decreases when PWM is more than 200, which is caused of energy consumption. As we can see from Figure 3 that the temperature of the top of the rack increases when tilted active tile angle is below 30 • . Because airflow does not reach the top of the rack. When the TAT angle is 35 • , the airflow blows over the racks. According to Figure 3 the performance of the tilted active tile is the best at 30 • angle. The temperature of the rack reduced under 30 • angle with small PWM. Temperature distribution is also a biggest problem in data centre, so using the tilted active tile to make the rack inlet temperature distribution uniform, clearly observed from the Figure 4 that when we adjust TATs at 30 • angle, and increase the PWM from 150 the temperature distribution of three position of the rack is more uniform. The temperature of the upper, middle and lower points of the rack was reduced from 19.2-20 to 17.35-17.7, best at 30 • angle. As we know that the empty As we know the empty U part of the FIGURE 3 Reduced temperature at different PWM rack cause the hot air circulation. So that's why we close the gap at bottom of the rack with black panel and repeated the experiment at TATs 30 degree angle. When the TAT Angle is 30, it can not only reduce the inlet temperature of the rack but also provide uniform temperature. Moreover, when the TAT was at 30 • angle, and increased PWM from 150, then the temperature distribution of the different position of the racks was more uniform.

Thermal performance analysis of rack gap sealing
We also analysed that the gap at the bottom of the rack and the empty U part of the rack also cause the hot air circulation. We analysed that the performance of the TATs U part of the rack causes the hot air circulation. So that is why we closed the gap at bottom of the rack with black panel and repeated the experiment at TATs 30 • angle. When the TAT angle is 30 • , it can not only reduce the inlet temperature of the rack but also provide uniform temperature.
We also analysed the rack thermal performance without sealing the gap at bottom of the rack. The D-04 rack is selected for analysis. Under all fan speeds, the inlet temperature of the rack with PWM of 0, 50, 100, 150, 200 and 250 was selected for analysis. The temperature distribution of the three temperature points on the rack was shown in (see Figure 5) at different angles. Figure 5 shows the average temperature at different angle. If the rack temperature is greater than 0 then it means the temperature of the rack is decreasing and if the temperature is lower than 0 it means temperature of the rack is increasing, which is caused of hot spot. When we adjust tilted active tile at 30 • angle, the average temperature of the rack increases at the bottom, middle and top of the rack which means sealing the gap not just decreases the rack temperature but also prevents the hot air circulation and improves the thermal performance of the rack. When we adjust tilted active tile except 30 • angle, so the average temperature of the bottom, middle and top of the racks are lower, which means except 30 • angle rack thermal performance is not good with gap sealing. Figure 6 shows the temperature distribution of the rack at the upper, middle and bottom points with different PWM. As we can see in Figure 6 that when PWM increases from 150 the temperature distribution of the rack at different positions are more uniform and the temperature at the three points of the rack is gradually decreased. The reason is that the gap of the rack is sealed so that is why the inlet temperature of the rack is more uniform, and the temperature of the rack is mainly affected by the tilted active tile. The temperature range of the upper, middle and lower three temperature points on the rack was reduced from 19.2-20 to 17.35-17.7. When the tilted active tile angle is 30 • , it can not only reduce the inlet temperature of the rack, but also make the temperature distribution of the three points of the rack more uniform. Figure 7 shows the average temperature of the rack and the temperature change in different PWM. We can see clearly in the picture that the rack temperature error gradually reduces with the increase of PWM.
Conclusion: In this section, we analysed the rack performance with sealing empty U part of the rack. It can be seen in the figure that the performance of the TATs is best at 30 • angle, which can make the rack temperature significantly lower compared with the passive tile. The temperature decreased at the bottom of the rack is the maximum under the condition of no gap sealing. While after sealing the gap of the rack can reduce the temperature about 3°C. When the gap of the rack is not sealed, the temperature difference between the three temperature points of the rack with the minimum overall variance of the rack is 1.0°C, while the temperature difference between the three temperature points of the rack with the gap sealed is only 0.179°C. We analysed that sealing the gap of the rack can reduce the hot air circulation.

DISTURBANCE ANALYSIS OF TILTED ACTIVE TILE
The deployment of the tilted active tiles in the cold aisle environment affects the air flow of surrounding racks, changing the direction of air flow can disturb the inlet temperature of the surrounding racks, so it is necessary to analyse the disturbance of the TATs. In this section we analysed the impact of TATs to the surrounding racks and the disturbance between the TATs.

Influence of tilted active tile on adjacent racks
The main purpose of the TATs is to solve hotspot problems of multiple racks in the data centre, therefore it is necessary to analyse the influence of TATs on adjacent rack in cold aisle environment. Figure 8 shows the reduced average temperature of the rack at three different positions of the rack under different PWM. As we can see in Figure 8, TATs have positive impact on adjacent rack. In this experiment, we adjust TATs at 0 • and 30 • angle and then evaluate the TATs impact on adjacent rack. As we can see in Figure 8 the TATs at 30 • angle has positive impact on surrounding rack. The average temperature decreased more when the TATs angle is 30 • . When the TAT angle is 0 • , the temperature decrease of the rack is not as obvious as the angle of 30 • , so the TATs at 30 • angle have positive impact on adjacent rack.
As we can see in Figure 9 when we adjust TATs at 30 • angle, the temperature of the bottom, middle and top of the rack change. According to Figure 9 the three temperature points of the rack under small PMW are more uniform. The temperature fluctuations of the rack is small under small PWM and with the increase of PWM, three temperature point of the rack is not more uniform. The reason is when the fan speed of the TATs is slow, the temperature of the rack is mainly affected by cold air blow from their surrounding rack. Figure 10 shows the overall mean temperature of the rack and the error change of the overall temperature of the rack. It can be seen from the figure that when PWM is 40, the overall temperature error of the rack reaches the minimum. The maximum temperature difference between the middle and lower temperature points on the rack is 0.429 • C.

Disturbance analysis of adjacent tilted active tiles
In this section we analyse the impact of adjacent TATs on the surrounding rack. We deploy two TATs next to each other, then analyse the impact of adjacent TATs on surrounding rack. As we can see in (Figure 11) that the rack inlet temperature of D-04 rack cannot change at constant fan speed of TATs but the influence of adjacent tilted active tile under D-05 rack, change the average rack inlet temperature, where the solid line represents the average rack inlet temperature at different PWM and the dotted line represents the average rack inlet temperature at different PWMs on adjacent tilted active tilted.
As we can see in the figure with the increase of fan speed the average temperature of the rack reduced. When the PWM of the TATs is 150, the temperature at the top of the rack near to the reference temperature. When the PWM of the TATs is 250, the temperature at the top, middle and lower points of the rack is greater than the reference temperature. The variation of the three temperature points of the rack is not consistent under different PWM.

Disturbance of a passive tile between two tilted active tiles
In this experiment we deploy a passive tile between two tilted active tiles and then analyse the thermal performance of surrounding racks. Figure 12 describes that the rack inlet temperature cannot change at constant fan speed of active tile. As we can see in Figure 12, when PWM is 50, the temperature of the middle of the rack drop the most relative to the reference temperature. The maximum temperature difference is 2.5 • C. When

Disturbance of two passive tiles between two tilted active tiles
In this experiment we deployed two passive tiles between two TATs and then analysed the thermal performance of the surrounding racks. Figure 13 shows the average temperature of the D-04 rack under the influence of the TATs. As we can see in the figure that the average temperature of the rack under TATs at PWM 150 and 250 appear below from reference tempera-ture, the rest of the PWM the average temperature appear above from the baseline temperature, the reason is slow fan speed of the TATs under the D-07 rack. However, the rack D-04 inlet temperature is mainly affected by its surrounding TATs.
Conclusion: As we know the thermal performance of the rack is better when we adjust TATs at 30 • angle. So using 30 • angle of the TATs to analyse the impact of disturbance TATs on surrounding rack. We analysed the disturbance of the TATs and passive tile, (1) TAT at 30 • angle has positive impact on the surrounding rack, but with increasing of distance from the TATs to rack, weaken the impact; (2) the closer the distance between tilted active tile and rack and increase the fan speed has negative impact on surrounding rack.

THERMAL MODELLING USING ARTIFICIAL NEURAL NETWORK
ANN-based thermal models to predict the thermal performance of TATs. The prediction accuracy of the model was extensively compared and analysed, i.e. BP and LSTM, were evaluated. LSTM provide short term memory into long term memory block. LSTM use for large amount of data and remembering information for long time. In BP, initial system output compared to the desired output and system is adjust until the difference between two minimized. We built multi tilted active tiles performance model. In these model we analysed the single output model and multiple output model. Both BP and LSTM models are divided into two types: single output model and multiple output model.

Single output model
The single output model can predict the temperature at different heights on the rack. As we can seen in Figure 14, the input of the single output model is wind temperature, fan speed and the height of the sensor. The output of the single output model is sensor height. BP and LSTM single output model as defined as BP single and LSTM single.

Multi-output model
The multi-output model can only predict the temperature at a fixed height of the rack. As we can seen in Figure 15, the input of the multi-output model is the outlet air temperature, fan speed and sensor height, and the output is the temperature at the fixed height of the rack. BP multi output and LSTM multi-output model are defined as BP many and LSTM many.

Single tilted active tile model optimization
In this experiment we adjust single active tile in cold aisle environment and then we analysed the thermal performance of the racks. Thermal efficiency model was established by BP and

Single output model optimization
In the experiment of a single tilting active tile, first analyse the single output model. In single output model BP performance  Figure 16) shows the average prediction errors of BP and LSTM models under different number of neuron in the training set. We train the model 36 times with different random seeds and then we evaluate the performance of model. The main purpose of this work is to analyse the prediction error and evaluate the performance of the model. As we can see in Figure 16, we train the model on the whole dataset and then observe that the prediction error of BP is minimized at 52 number of neurons and prediction error of LSTM is minimized at 88 number of neurons. It can be seen from Figure 16 that the average prediction error of BP neural network is smaller than LSTM. Figure 17 shows the prediction error of the two models at six temperature points in the training set and the validation set. It can be seen from the figure that the prediction error of BP in the training set is smaller than LSTM. In the training set and validation set, the accuracy of BP and LSTM improved by 16.4% at the six temperature points of the whole racks. It can be seen from the figure that the average prediction error of BP at different temperature points is smaller than LSTM.
It can be seen from Figure 18 that the average prediction error of BP is smaller than LSTM. The prediction error of the

FIGURE 10
The overall average temperature and error of the D-05 rack at active tile angle of 30 • middle position of the rack is small. Because the temperature distribution at the bottom and top of the rack is uneven. As a result, the model cannot make accurate predictions based on the height of the sensor, there is a large temperature difference between the left side sensors of the rack and right side sensors, the difference between the sensors is only 15 cm in the vertical direction.
BP neural network was selected to continue the analysis, the error distribution of the model in the test set is shown in Figure 19. It can be seen from the figure that only t10 and t9 temperature points have most data errors between 0 and 0.43, However, t7 at the top of the rack, t8 at the bottom of the rack have large errors due to the large temperature difference between the temperature points on the rack, Therefore, the generalization of neural network is not valid for all temperature point.

5.3.2
Multi output model optimization Figure 20 shows the average prediction error of BP and LSTM in the training set with different number of neurons. The main purpose of this experiment is to analyse the prediction error and evaluate the performance of the model. We train the whole dataset and then observed that the prediction error of BP is minimized at 96 number of neurons and prediction error of LSTM is minimized at 84 number of neurons. It can be seen from Figure 20 that the average prediction error of LSTM neural network is smaller than BP, because the minimum loss error of LSTM is 0.0403 at 84 number of neurons and BP loss error is 0.0922 at 96 number of neurons which mean LSTM performance is better than BP. Figure 21 shows the predicted mean errors of the two models in the test set, it can be seen from the figure that the average prediction error of LSTM model in t2-t6 is less than 0.09. The model has high prediction accuracy. The prediction errors of LSTM model in t4 and t5 are the most different from BP. It decreased by 100.93% and 92.99% respectively.
The average error of the two models in the training set and validation set is shown in Figure 22. According to Figure 22 the average error of LSTM is smaller than BP.  analysis and comparison, it is known that the temperature distribution on the rack is not uniform. In the prediction accuracy experiment of the model, the LSTM model has strong prediction ability, the prediction error of the model is basically between 0 and 0.14.

5.4
Two tilted active tile model optimization As we can seen in Figure 24, we adjust two TATs next to each other in cold aisle environment. We analysed the thermal performance of rack. Thermal efficiency model was establish by BP

FIGURE 17
Model performance on different data sets

FIGURE 18
Error histogram for different model

5.4.1
Single output model optimization Figure 25 shows the average prediction errors of BP and LSTM models under different numbers of neuron in the training set. It can be seen from the figure that the average prediction error of BP neural network is smaller than LSTM. As we can see in Figure 25 that the minimum loss error of BP is 0.0646 at 88 number of neurons and LSTM loss error is 0.1952 at 56 number of neurons, which means BP performance is better than LSTM, because the error of BP is minimum than LSTM. Figure 26 shows the prediction error of the two models at six temperature points in the training set and the validation set. It can be seen from the figure that the prediction error of BP in the training set is smaller than that of LSTM. In the training set and validation set, the accuracy of BP and LSTM improved by  106.28% and 104.59% respectively at the six temperature points of the whole racks. Figure 27 shows the average prediction error of BP and LSTM neural network models. It can be seen from the figure that the prediction error of LSTM is greater than that of BP neural network. In terms of the overall average prediction error of the rack, BP is 11.75% smaller than LSTM. The prediction accuracy of BP model is better than LSTM.
According to Figure 28, BP neural network is selected to continue the analysis. The error distribution of its model in the test set is shown in figure. It can be seen from the figure that only the temperature point t11 has a small error range, while the error fluctuation range of other temperature points is large, among which t9 is the most obvious. It can be concluded that the gen-

5.4.2
Multi output model optimization Figure 29 mainly shows the average prediction error of BP and LSTM training sets under different number of neurons. It can be seen from the figure that the average prediction error of LSTM neural network is smaller than BP, because the minimum loss error of LSTM is 0.0718 at 84 number of neurons and BP loss error is 0.0796 at 44 number of neurons which mean LSTM performance is better than BP, because the error of LSTM is minimum than BP. Figure 30 shows the prediction error of the two models at 6 temperature points in the training set and the validation set. It can be seen from the figure that the prediction error of LSTM in the training set is smallerthan that of BP. Figure 31 shows the prediction error of BP and LSTM models in the test set. It can be seen from the figure that the average prediction error of LSTM is smaller than BP except t t1 temperature point. The LSTM model has the best prediction accuracy. Figure 32 shows the error distribution of LSTM neural network in the test set. It can be seen from the figure that the data with errors between 0-0.05 and 0.05-0.1 are the most, and the data with errors greater than 0.24 are very small. The average

Two tilted active tile between one passive tile model optimization
In this experiment we adjust a passive tile between two active tiles in cold aisle environment. We analysed the thermal performance of rack. Thermal efficiency model was establish by BP and LSTM. In this experiment single and multi-output model were analysed. Single output model optimization Figure 33 shows the average prediction errors of BP and LSTM models under different number of neurons in the training set. We train the whole data set and analysed the minimum average error of BP and LSTM and then compared the performance of these two models. It can be seen from the figure that the average prediction error of BP neural network is smaller than LSTM, and the error of BP is 67.87% lower than LSTM. It can be seen from Figure 34 that the average prediction error of BP and LSTM reached the minimum point when number of neurons reached 64 and 48 respectively. Figure 35 shows the prediction error of the two selected models at six temperature points in the training set and the validation set. It can be seen from the figure that the prediction error of BP in the training set is smaller than LSTM. In the training set and validation set, the accuracy of BP was 211.22% and 199.76% higher than that of LSTM respectively at the six temperature points of the rack.   Figure 36 shows the average error of BP and LSTM neural network models with the least error in the training set to predict the data in the test set. As we can see from the figure, except t11 and t7, the prediction error of the other temperature points of LSTM is greater than that of BP neural network. In terms of the overall average prediction error of the rack, BP is 3.70% lower than LSTM. The prediction accuracy of BP model is high than LSTM.
According to Figure 37, BP neural network is selected to continue the analysis. The error distribution of its model in the test  Figure 38. It can be seen from the figure that only the temperature point t11 has a small error range, while the error fluctuation range of other temperature points is large, among which t9 is the most obvious The results are consistent with those predicted by the single output structure model of adjacent active floor. It can be concluded that the generalization of neural network is not effective for all temperature points, and the prediction result is poor.

5.5.2
Multi output model optimization Figure 38 mainly shows the average prediction error of BP and LSTM in the training set under different numbers of neuron. It can be seen from the figure that the prediction error of BP and LSTM models reaches the minimum when the number of neurons is 44 and 88 respectively, and the average prediction error of LSTM in the training set is less than BP. Figure 39 shows the prediction error of the two models at 6 temperature points in the training set and the validation set. It can be seen from the figure that the prediction error of LSTM in the training set is smaller than BP, but in the validation set, the prediction error of LSTM in t5 is 16.5% higher  Error histograms of different models in the test set than that of BP. In the training set and validation set, the LSTM accuracy improved by 28.89% and 13.78% respectively compared with BP at the six overall temperature points of the rack. Figure 40 shows the prediction errors of BP and LSTM models in the test set. It can be seen from the figure that the prediction errors of LSTM at different temperature points are smaller than BP. Figure 41 shows the error distribution of LSTM neural network in the test set. As we can see from the figure, the data errors are basically distributed between 0 and 0.64. The maximum prediction error of other temperature points is less than 0.32, and the maximum prediction error of t5 is less than 0.45. LSTM has high prediction accuracy.

5.6
Two tilted active tile between two passive tiles model optimization As we can seen in Figure 42, we adjust two passive tile between two active tiles in cold aisle environment. We analysed the thermal performance of rack. Thermal efficiency model was establish by BP and LSTM. In this experiment single and multioutput model were analysed.

5.6.1
Single output model optimization Figure 43 shows the average prediction error of BP and LSTM models under different number of neurons in the training set. It can be seen from the figure that the average prediction error of BP neural network is smaller than that of LSTM, and the error of BP is 66.53% smaller than that of LSTM. It can be seen from the figure that the prediction error of BP and LSTM models reaches the minimum when the number of neurons is 68 and Error histograms for different models 76 respectively, and the average prediction error of BP in the training set is less than LSTM. Figure 44 shows the prediction error of the two models at six temperature points in the training set and the validation set. It can be seen from the figure that the prediction error of BP in the training set is smaller than that of LSTM. In the training set and validation set, the accuracy of BP increased 198.86% and 183.84% respectively compared with LSTM at the overall six temperature points of the racks. Figure 45 shows the average error of BP and LSTM neural network models in the training set. It can be seen from the figure that the prediction error of BP model is greater than LSTM, except t8, the prediction error of BP model is the smallest. The average prediction error of BP was smaller than that of LSTM, which decreased by 1.03%.
According to Figure 46, BP neural network is selected to continue the analysis. The error distribution of its model in the test

5.6.2
Multi output model optimization Figure 47 mainly shows the average prediction error of BP and LSTM in training sets with different number of neurons. It can be seen from the figure that the prediction error of BP and LSTM models reaches the minimum when the number of neurons is 32 and 72 respectively, and the average prediction error of the LSTM is smaller than BP model. Figure 48 shows the prediction error of the two models at six temperature points in the training set and the validation set. It  can be seen from the figure that the prediction error of LSTM in the training set is smaller than that of BP. In the training set and validation set, the accuracy of LSTM increased 198.86% and 183.84% respectively. Figure 49 shows the prediction error of BP and LSTM models in the test set. It can be seen from the figure that the prediction error of LSTM is smaller than BP, except t1. The prediction accuracy of LSTM neural network is higher than that of BP neural network. Figure 50 shows the error distribution of LSTM neural network in the test set. It can be seen from the figure that the data errors are basically distributed between 0 and 0.2, except t4 and t5. The maximum prediction error of other temperature points is less than 0.45, and the maximum prediction error of t4 and t5 is close to 0.7. The prediction error of most data is close to 0, which further indicates that the model has good prediction ability.

DISCUSSIONS ON TATS IMPACT ON RELIABILITY
Data centre is a large scale computing system which requires extremely high level of reliability. TATs not only improve the cooling air delivery, but also enhance the system reliability. The main purpose of this study was to determine if proposed TAT provide any advantage over passive tiles in terms of increasing ride through period for a cooling disruption. A cooling failure in a data centre can occur due to a power, or mechanical component failure. In this study we discuss the case when CRAC blower fails. The proposed TAT has a longer ride through time for scenario CRAC blower failure. In case of CRAC blower failure, the TAT fans are still running and there is some recirculation of air which can absorb heat from the IT equipment. For example in case of cooling failure in data centre TAT can provide cooling up to 10 min and passive tile stop cooling on the spot. Possible reason for this is growth in thermal mass available in the room and/or air circulation. In particular tiles fans can partially drive the airflow, thus improve the thermal condition even if the CRAC blower malfunction [21].

CONCLUSION
The main purpose of this paper is to reduce the energy consumption in data centre and improve the resilience of data centre cooling system by using TATs, which are designed to improve the local cold air supply. TATs also prevent the air flow blow over the rack and eliminate the hot spot. TATs is used to prevent the hot air recirculation and reduced the energy consumption of cooling system. In case of power failure in data centre, the proposed approach can maintain the rack inlet temperature for up to 10 min. An extended ride-through time was observed for data centres deployed with TATs when the CRAC blower failed. We also analysed the thermal performance of the TATs. We adjusted TATs at different angle, and compared the thermal performance of TATs under different environment. We adjusted multiple TATs at different positions in the cold aisle and then analysed the thermal performance of the racks. It was concluded that the best thermal performance is achieved when TATs are at 30 • angle. We also introduced ANN-based thermal models to predict the thermal performance of the racks. Thermal efficiency model was established by BP and LSTM, in this experiment single output model and multi output model were analysed. The single output model can predict the temperature at different heights on the rack. In single output model the predicted effect of BP model is better than LSTM. The average prediction error is 0.57. The multi-output model can only

FIGURE 50
Prediction error distribution for test set predict the temperature at a fixed height of the rack. In multi output model LSTM model is better than BP. LSTM prediction error is less than BP. The average prediction error is 0.07. In this experiment we also observed that the empty server slots in the rack cause hot air circulation. After sealing the gap of the rack, the performance of the TATs can be significantly improved.
Our experimental results also support that the airflow of TATs has positive influence on surrounding racks in terms of cooling performance.