Daily average relative humidity forecasting with LSTM neural network and ANFIS approaches

Because hurricanes, droughts, floods, and heat waves are all important factors in measuring environmental changes, they can all result from changes in atmospheric air temperature and relative humidity (RH). Besides, climate, weather, industry, human health, and plant growth are all affected by RH. Accurately and consistently forecasting RH is a challenge due to its non-linear nature. The present study tried to predict one day ahead of RH in determined provinces from different climatic regions of Turkey (Ankara, Erzurum, Samsun, Diyarbakır, Antalya, and Bilecik) using long short-term memory (LSTM) and adaptive neuro-fuzzy inference system (ANFIS) with fuzzy c-means (FCM)–based machine learning models. As evaluation criteria, root mean square error (RMSE), mean absolute error (MAE), and correlation coefficient (R) were employed. The outcomes from the forecasting models were also validated using observed data. During the testing stage, the smallest MAE and RMSE values were discovered to be 5.76% and 7.51%, respectively, in Erzurum province, with an R value of 0.892 when using the LSTM method. Moreover, the smallest MAE and RMSE values were obtained to be 5.95% and 7.67%, respectively, in Erzurum province with an R value of 0.887 using the ANFIS model according to the daily RH prediction. The results indicate that both the LSTM and ANFIS approaches performed satisfactory performance in daily RH prediction, with the LSTM and ANFIS approaches producing nearly identical results.


Introduction
Relative humidity (RH) is the ratio of the amount of water vapor available in the air to the maximum amount of water vapor that can be held in the air at the same temperature. In other words, RH is defined as the ratio of moisture partial pressure in the air to water vapor saturation pressure at the same temperature. If the RH of the air at a given temperature is 100%, it indicates that the air has reached saturation, and if more moisture is added, condensation will form in the air. In addition to humans, plants, animals, and even inanimate objects are affected by the RH in the air. At a certain temperature and humidity, people feel very comfortable both indoors and outdoors. Efforts are being made to provide the necessary comfort conditions using a variety of methods. RH is critical in providing the necessary comfort, particularly for livestock, as well as for human comfort. To maximize the efficiency of livestock, the RH of the environment in which the animals are housed must be kept under control (Zhou et al. 2021). RH governs evaporation at the earth's surface as well as plant leaf transpiration. Furthermore, it prevents solar radiation from reaching the earth, thereby preventing overheating or cooling. The importance of RH in food storage cannot be overstated. To avoid food damage, precautions must be taken depending on whether the RH is low or high. Furthermore, RH affects roads, buildings, machinery, electrical transmission lines, and various mechanical and electronic devices and tools we use. All of these can be harmed by both low and high RH. RH must be kept under control at all times during the manufacturing and use processes. As a result, predicting RH is critical for design and manufacturing processes.
The study conducted by Parishwad et al. emphasized that RH, air temperature, and wind speed were the most important parameters in the sizing of air conditioning systems and that optimum sizing could be accomplished by accurately estimating these parameters. Using meteorological data from various cities in India, a model was developed to estimate RH, air temperature, and wind speed. The error rates for air temperature, RH, and wind speed in the developed model were reported to be 10.5%, 14.6%, and 26.7%, respectively (Parishwad et al. 1998). Mass and energy balance equations are utilized in the physical modeling of a modern building for automatic control of the HVAC system. To obtain accurate models using this method, the physical properties of the building must be well understood. Inadequate information can have an impact on HVAC facility design and control, resulting in increased energy consumption and poor indoor air quality. The artificial neural network model proposed by Mustafaraj et al. was used to predict RH and indoor air temperature. For a set period of time, indoor and outdoor temperature and RH values were recorded, and the resulting data was used in the artificial neural network model. Very accurate results can be obtained, and complex problems can be solved much more easily and quickly using this method, which does not require as detailed data entry as physical models (Mustafaraj et al. 2011). Because of rising energy costs and current environmental regulations, it is possible to argue that dynamic control is required for HVAC systems. The RH value, as well as other meteorological data, is required for the dynamic control of HVAC systems. Significant energy savings will be possible thanks to the estimation of meteorological data using various methods and the dynamic control of HVAC systems (Mustafaraj et al. 2010 Although there are many studies for predicting temperature, it has been noted that there are only a few studies for predicting RH. It has been stated that by predicting a city's RH value, the energy management systems of the buildings can be designed accordingly, resulting in energy savings in the buildings (Martin et al. 2015). Also, RH and temperature have a substantial effect on the performance of air-source heat pumps operating in different climatic conditions. Icing in the heat pump may occur depending on the RH value in the outdoor environment, resulting in performance loss. Knowing how RH varies by climatic region and taking appropriate measures will also help to prevent heat pump performance loss (Wu et al. 2020).
There are other areas where the RH is effective. For example, high RH levels combined with high air temperatures can cause more thermal stress and death in humans. According to research, low humidity levels play a significant role in severe influenza epidemics, particularly in autumn and winter in temperate climates. For these reasons, the significance of forecasting humidity fluctuations and extremes and taking appropriate measures was emphasized (Langendijk et al. 2021). Liu et al. looked at how temperature and RH affected learning performance. It was determined that RH is a more important factor in learning performance than temperature ). In the study conducted by Pei et al., it was investigated how temperature and RH affect the filters used to provide the necessary air standards for laboratories. In accordance with the findings of the study, RH, rather than temperature, influences the performance of the filters (Pei et al. 2021). Also, RH can affect the evaporation rate and equilibrium size of particles in the air, which can affect particle removal rates and virus viability.
The amount of energy consumed in buildings is affected by indoor thermal comfort conditions. It is generally determined by the RH and air temperature. For the estimation of comfort parameters, simulation programs have been developed. However, in order to use them, extremely detailed data must be provided. According to Mba et al., predicting temperature and RH values and transferring this information to the building energy management system can save a significant amount of energy (Mba et al. 2016). Using building energy simulation tools is an efficient way to assess indoor air temperature and RH parameters. To make the simulation tools work properly, all kinds of detailed parameters about the building must be collected, such as changes in weather conditions, the structure of the building, its geographical location, the lights, and the energy generated by the equipment. It takes time and a lot of effort to collect these parameters. Most of the time, accessing all of this data may be impossible. Furthermore, the development of these simulation models necessitates expert knowledge, and the computation process is time-consuming. On the other hand, it is well known that some nonlinear models, such as artificial neural networks, outperform linear models. Because neural networks can infinitely approach any continuous function, they are widely utilized in various fields. Although many temperature prediction models have been developed, studies on RH are limited (Shi et al. 2018).
The thermal conductivity of building components is an important parameter that directly affects building heating and cooling loads. The effects of RH and air temperature on the thermal conductivity of building insulation materials have been studied in the literature. The thermal conductivity of insulation materials is measured at room temperature and in dry conditions. In practice, building walls are subjected to changing outdoor conditions such as RH and air temperature. When the temperature rises from 20 to 60 °C, the heat transmission coefficients of the insulation materials used in buildings rise by 8.8% to 21.4%, and the thermal conductivity coefficient rises by 14.4% when the RH rises from 0 to 100%. It was discovered that there was an increase ranging from 14.8 to 186.7%. As a result, the amount of heat transferred from the building walls was estimated to have increased by 1.5 to 1.9 times (Wang et al. 2022). In a study conducted by Karyono et al., it was discovered that estimating the RH value and using this data in the building energy management system could save 2% of the energy consumed in buildings (Karyono et al. 2022).
In a study, Csavina et al. demonstrated the effects of RH and wind speed on atmospheric dust concentrations in semiarid climates. It has been emphasized that estimating the harmful effects of dust and aerosol emissions in the air is possible and that estimation can be performed with the help of wind speed and RH parameters (Csavina et al. 2014). Humidity is a significant component of the hydrological cycle, influencing both the air and the climate. The concept of RH is widely applied in many fields, including hydrology, ecology, agriculture, and medicine (Gunawardhana et al. 2017). Temperature and RH both have a significant impact on air quality and human health. As a result, mapping RH and temperature extremes during the hot summer months is critical for public service and environmental policy (Li and Zha 2018).
The RH of the air must be taken into consideration in the food and agricultural industries as well. The effect of RH on water loss during pear storage was studied. The finite element method was used to analyze a three-dimensional model for this. Increasing the RH from 95 to 97% reduces water loss in pears by 3%, according to the findings. As a result, the significance of controlling indoor RH by estimating RH change during storage was emphasized (Nguyen et al. 2007). Biaobrzewski's study highlighted the importance of RH in grain storage. The artificial neural network method has been claimed to be precise in estimating RH (Białobrzewski 2008). Because the temperature and humidity controls have a direct impact on product quality, it is critical in industries such as tobacco processing, food production, and biological product production (Shi et al. 2018). Kim et al. studied the effects of RH on egg quality and laying hen physiological stress. Based on what we learned from the research, high RH reduces egg quality and increases stress in chickens (Kim et al. 2021). The importance of RH during mushroom drying was emphasized in a study conducted by Li et al. (2022).
The importance of RH and air temperature control in the production of electronic devices was emphasized by Conseil-Gudla et al. It has been stated that condensation may occur as a result of humidity fluctuations, which will have a negative impact on electronic devices. It has been stated that the necessary precautions must be taken by predicting the RH (Conseil-Gudla et al. 2018). According to Shi et al., accurate, reliable, and stable RH and air temperature controls in industrial production are critical for efficiency. Simultaneously, the importance of accurately predicting air temperature and RH in climate control systems in terms of system stability was emphasized. The artificial neural network method was shown in the study to be very sensitive for estimating RH (Shi et al. 2018). Sohani et al. studied the effects of absolute and RH on photovoltaic performance using the artificial neural network method. It has been discovered that as absolute and RH levels rise, so does the performance of photovoltaics (Sohani et al. 2020). The relationship between RH and metallurgical engineering is also very close. The effect of RH on the production process of aluminum alloys was highlighted in the study conducted by Safyari et al. As a result of the research, cracks can form in aluminum alloys when RH is high, and RH should be kept under control during the manufacturing process (Safyari et al. 2021). Temperature and RH effects on high voltage power transmission lines were investigated. It has been stated that high voltage line initial investment costs can be reduced by taking the necessary measures in high voltage lines based on seasonal and regional changes by predicting the RH value .
In another investigation, the researchers indicate that the curing times and strengths of concrete are determined based on the assumption that the RH value is 100%. However, in field studies, this is frequently not possible to provide, and the concrete strength values calculated based on the 100% RH are higher. Indeed, it has been emphasized that the RH value of the outdoor environment shapes concrete strengths, so the RH value should also be taken into account in field studies (Liao et al. 2008). As a result, it is critical to know the RH value of the environment before pouring concrete and to take appropriate precautions in concrete construction projects that are exposed to the outside environment. Kwon et al. proposed a method for estimating the compressive strength of concrete used in the construction of nuclear power plants. The study revealed that, in addition to temperature, RH is an important parameter in concrete strength, and the importance of predicting air temperature and RH parameters for the concrete strength of nuclear power plants to be built in different climatic regions and seasons was emphasized (Kwon et al. 2014). Shen et al. investigated the impact of RH on the curing time of concrete and crack formation in concrete (Shen et al. 2017). Temperature and humidity are known to cause corrosion in reinforced concrete structures. Chauhan and Sharma investigated the effects of air temperature and RH changes on corrosion in reinforced concrete, concluding that both parameters should be estimated for each climate zone and reinforced concrete reinforcement calculations should be made accordingly (Chauhan and Sharma 2019). Xi et al. studied the effect of RH on the linear viscoelastic properties of asphalt mixtures (Xi et al. 2021).
Interpolation, simulation, and regression approaches are commonly applied to predict RH and air temperature. The interpolation method is the simplest approach, but it introduces significant errors and uncertainties (Vincent and Mekis 2006;Yang et al. 2004). For more precise estimations, regression methods can be used (Li and Zha 2018). Gunawardhana et al. proposed an alternative model for estimating RH. Meteorological data from many years was statistically converted into a smaller data set in the proposed method, and a RH forecasting model was developed as a result. It was stated that the obtained results were very good (Gunawardhana et al. 2017). The artificial neural networks method was used in the study conducted by Kuzugudenli to estimate RH using meteorological data from 177 different points in Turkey. According to the results, the method used for estimating RH produced very sensitive (R 2 = 0.84) results (Kuzugudenli 2018). Accurate outdoor temperature and RH estimation have a crucial impact on agricultural and industrial production, as well as environmental and economic policies. It is possible to predict the outdoor air temperature and RH very precisely using models developed using the artificial neural network method (Al-Shawwa et al. 2018). Hutapea et al. conducted research on the estimation of RH using the LSTM network method (Hutapea et al. 2020). Adnan et al. emphasized the importance of accurate and reliable estimation of RH in all areas related to global climate change. A machine learning method was used to estimate RH in the Hunza River basin in Pakistan. It has been determined that the error rates in the results obtained using the method are quite low, and thus, the multivariate adaptive regression spline (MARS) method can be easily used in the estimation of RH (Adnan et al. 2021). Due to its highly complex and non-linear nature, research on RH estimation is limited. In the study by Qadeer et al., a machine learning method was used to estimate RH. In accordance with reports, the machine learning method can predict RH very precisely, and this data can be used in various applications, including the operation and design of HVAC systems, cooling towers, and thermal power plants (Qadeer et al. 2021).

Research significance and novelty of the study
As can be seen in the literature research described above, the effects of RH on various fields, such as humans, animals, foodstuffs, air-conditioning applications, electronic devices, and building materials, have been investigated, but there have been few studies on its forecasting using a machine learning approach. Because of this, it is anticipated that this study will fill a gap in the literature and make major contributions to the field.

Study area and data analysis
The calculations were performed using the average daily RH values measured by the General Directorate of State Meteorology Affairs between 2012 and 2019 in 6 different provinces (Ankara, Erzurum, Samsun, Diyarbakır, Antalya, and Bilecik) selected from 6 different climatic regions of Turkey as shown in Fig. 1. Table 2 gives the geographic coordinates of the stations as well as a statistical summary of the dataset. As seen from the table, the Diyarbakır station has the highest RH fluctuation and the lowest average RH value. Furthermore, Samsun station has the smallest standard deviation value of 12.31%.
The main steps of the suggested machine learning techniques shown in Fig. 2 can be expressed as follows (Adnan et al. 2021): 1. The data was pre-processed and revised for the next step by interpolating missing values after it was collected from the stations.
2. The original data were then normalized and the proposed models trained, validated, and tested, respectively. 3. Each model was applied to these data separately with the same input data combinations for RH prediction. 4. To examine the precision of the developed model, various statistical criteria such as MAE, RMSE, and R were used. 5. Eventually, RH was forecasted using the designed machine-learning approaches.
MATLAB R2018a (Trail Version) was used to perform the models.

Long-short term memory (LSTM) neural network
A long short-term memory (LSTM) neural network is expressed as a type of recurrent neural network (RNN) that could learn long-term dependencies between time steps of series data. Unlike RNNs, machine learning techniques based on LSTM are developed to prevent the long-term dependency problem. The LSTM neural network performs the training process using back propagation over time, thus eliminating the vanishing gradient problem. Conventional artificial neural networks have neurons, whereas LSTM neural networks have memory blocks connected via sequential layers (Abdel-Nasser and Mahmoud 2019). The basic elements of an LSTM neural network consist of a set of the input layer, an LSTM layer, a fully connected layer, and a regression output layer. This network structure functions first with input layers and then with an LSTM layer. With a series of the input layer, sequence or time-series data is entered into the network architecture. An LSTM layer gets to know long-term dependencies between time steps of sequence data. Next, the network completes the function with regression and fully connected output layers Ozbek et al. 2021;Sekertekin et al. 2021). Figure 3a shows an LSTM layer structure. This graph shows the flow of a time series X with C channels of length S through an LSTM layer. In the graph, c t and h t indicate the cell state and the hidden state at time step t, respectively. The first LSTM block utilizes the initial state of the network and the first time step of the sequence to calculate the first output and the updated cell state. At time step t, the LSTM block utilizes the current state of the network (c t-1 , h t-1 ) and the next time step of the sequence to calculate the output and the updated cell state c t (Mathworks 2021;Zolfaghari and Golabi 2021). Figure 3b shows the data flow at time step t for the LSTM block. It is clear from the figure that the gates forget, update, and output the cell and hidden states. The state of a layer is composed of the hidden state or output state and the cell state. The output of the LSTM layer is derived from the hidden state at this time step t. The cell state includes knowledge learned from the previous time steps. The layer adds knowledge to or extracts knowledge from the cell state at each step. The layer manages these updates using the gates. As seen in the LSTM layer architecture, input gate (i), forget gate (f), cell candidate (g), and output gate (o) components are included. These elements check the hidden state and cell state of the layer. The input gate checks the level of cell state update. The forget gate checks the level of cell state reset (forget). The cell candidate   (Hochreiter and Schmidhuber 1997). The input weights W, the recurrent weights R, and the bias b form the learnable weights of an LSTM layer. The matrices W, R, and b can be given as follows (Mathworks 2021): The cell state at time step t is expressed by where ⊙ demonstrates the element-wise multiplication of vectors. The hidden state at time step t is expressed by where c demonstrates the state activation function. The hyperbolic tangent function (tanh) is utilized to calculate the state activation function, and it is expressed by At time step t, the components of the input, forget, cell, and output gates can be identified as follows, respectively (Mathworks 2021): In these calculations, g demonstrates the gate activation function. The sigmoid function is utilized to calculate the gate activation function, and it is expressed by

Adaptive neuro-fuzzy inference system (ANFIS)
Although artificial neural networks (ANNs) are effective for modeling different natural processes, they also have their shortcomings (Tabari et al. 2012). The ANFIS, first Fig. 3 a LSTM layer architecture, b the flow of data at time step t presented by Jang (1993), is a universal predictive approach and thus can approximate any natural continuous procedure on a consolidated set with any grade of accuracy (Goyal et al. 2014). ANFIS, a Sugeno-type fuzzy systems network expression, is equipped with neural learning abilities. This technique operates similarly to neural networks and creates a method for fuzzy modeling procedures to learn information about a dataset. An ANFIS provides a strategy with the support of a fuzzy rule or rules. Fuzzy rules are statements that present the relationship between the inputs and outputs of the system in the form of if-then statements depending on linguistic variables (Ay and Kisi 2014).
It can be considered that the system of fuzzy inference consists of two x and y inputs and one z output. Assuming the rule base includes two fuzzy if-then rules of Takagi and Sugeno type: where A i and B i describe the fuzzy clusters in the antecedent, while r i , q i and p i depict the design factors in the training procedure. As can be seen in Fig. 4, the ANFIS structure consists of five layers.
Layer 1: This layer defines an input variable for every suitable fuzzy set. The nodes in this layer indicate the function by which membership degrees are formed using (10) Rule1 ∶ if x is A 1 and y is B 1 , then f 1 = p 1 x + q 1 y + r 1 (11) Rule2 ∶ if x is A 2 and y is B 2 , then f 2 = p 2 x + q 2 y + r 2 membership functions (MFs). The node function of a node i is described as (Jang 1993): where A i and B i are the MFs. Layer 2 In this layer, the product output is created by multiplying the incoming signals. All nodes determine the firing strength of a rule by multiplying the incoming signals.
Layer 3: The i th node in Layer 3 calculates the ratio of the i th rule's firing strength to the sum of all rules firing strengths of all rules: where w i is the normalized firing strengths. Layer 4: Node i in Layer 4 identifies how the i th rule will support the model output through the node function: where {r i , q i, and p i } is the parameter set, and w i is the output of the layer 3.
Fig. 4 ANFIS architecture structure: a type-3 fuzzy reasoning; b equivalent ANFIS (type-3 ANFIS) (Jang 1993) Layer 5: This layer is called the output nodes. Here, the single node computes the output considering the sum of all incoming signals (Tabari et al. 2012): Accordingly, the output z in Fig. 2 is found as

Error analysis
In our study, three statistical error criteria such as mean absolute error (MAE), root mean square error (RMSE), and correlation coefficient (R) are considered. Their calculation methods are as follows: where p(i) and o(i) describe the predicted and observed RH at the time i, respectively. Besides, p and o stand for the mean of the predicted RH and the observed RH, respectively. N represents the total number of RH data.

Fuzzy C-means clustering (ANFIS-FCM)
Initially, the ANFIS-FCM clustering model was chosen, and numerous clusters, such as epoch number, input number, and MFs were used in the simulation. The method's performance was then evaluated by computing MAE, RMSE, and R values. In the analysis, the influence of the epoch number, input number, and MF parameters were investigated. While the best results are obtained in Ankara, Erzurum, and Diyarbakir when 2 MFs and 5 input numbers are used against 10 epochs; it was obtained when 2 MFs and 5 input numbers were used against 100 epoch numbers in Samsun, Antalya, and Bilecik. It was seen that only the epoch number had an effect among the clusters.

Long short-term memory (LSTM)
Although 300 epoch numbers are used in all provinces with this method, the best results were obtained when 5 hidden layers were used in Samsun; 10 hidden layers in Erzurum, Diyarbakir, Antalya, and Bilecik; and 25 hidden layers in Ankara. In this method, while the effect of the hidden layer was observed, it was seen that the epoch number was not effective. As shown in Table 3, the lowest MAE value is equal to 5.76% and RMSE is 7.51% with an R value of 0.892 in Erzurum. In all analyses made for all provinces, it has been observed that increasing the number of hidden layers and epochs has no effect on decreasing the MAE and RMSE values. The highest MAE and RMSE values were found in Antalya province as 8.37% and 10.84%, respectively, with an R value of 0.726. Figure 5 depicts the RH values observed with the training (7 years) and testing (1 year) data for the provinces to be used in the ANFIS-FCM and LSTM algorithms. Two thousand nine hundred twenty daily observed data points were used in the training and testing processes. Within these measurement values, 2555 pieces of data (87.5% of the entire dataset) from 01 January 2012 to 31 December 2018 were used to train the models, and the last 365 daily observed data (12.5% of the entire dataset) from 01 January   In all provinces, the same training and testing ratios were applied. The X-axis in this figure displays the number of observed values, while the Y-axis represents the RH in %. As can be seen from the figure, the prediction of the RH time series by LSTM (red-colored dataset) corresponds to the actual values in the testing part in all provinces.
The results show that the ANFIS-FCM model and the LSTM neural network perform well with this type of sinusoidal data. To examine the prediction results in greater depth, Fig. 6 is a close-up of only the testing values. As can be seen from this figure, the estimations obtained by both ANFIS-FCM and LSTM methods in Diyarbakir are very close to the real values. While the most accurate estimates were obtained in Diyarbakir, the most remote values were obtained in Samsun.
The variation between the observed and predicted RH values is shown in Fig. 7. While the error value between observed and predicted RH data ranged between − 30 and + 30% in Ankara, Samsun, and Antalya provinces, the least error was obtained in the province of Diyarbakir. It is clear from the figures that, especially in the spring and summer months, error rates were determined to be much lower in Diyarbakir province. The reason for this is that the RH values are very low on these dates. Figure 7 also shows a plot of observed and predicted RH on a daily basis to reveal the data distribution, which supports the previous results. Moreover, the correlation coefficient (R) is presented in this figure. The observed and forecasted data obtained by the LSTM approach almost coincide with each other in Diyarbakir province with an R value of 0.9496. The smallest R value was obtained in Samsun with a 0.6082. Figure 8 shows histograms for all provinces that show the frequency distribution of RH differences between predicted and observed RH for ANFIS-FCM and LSTM methods. In particular, results obtained from Diyarbakır tend to show the closest thing to a normal distribution, with an almost symmetrical distribution of about zero. In that province, the error distribution follows a Gaussian curve. The results in Samsun and Antalya appear to be skewed to the positive, while the result in Ankara appears to be skewed to the negative for the models. In both the ANFIS-FCM and LSTM methods, more than 85% of the model predictions have an error ≤ ± 10% for Diyarbakır as can be seen in this figure. In Antalya, on the other hand, more than 72% of the model predictions have an error of ≤ ± 10%. For all the methods under consideration, greater than 90%, the error values are ≤ ± 16% and the high errors of about 24% and greater in prediction occurred only in Samsun and Antalya provinces.
The Taylor diagrams (Taylor 2001) in Fig. 9 were constructed for evaluating the results. As is known, a Taylor diagram allows for determining the accuracy of the proposed models in a variety of ways. The distance between the points representing the models and the point representing the observed value is inversely proportional to the overall similarity of the models to the measured values. In comparison to the ANFIS-FCM model, Fig. 9 clearly shows that the LSTM model predictions are closer to the observations. The lowest RMSE value of 7.51% was obtained in Erzurum when the LSTM method was used. In Diyarbakir, a correlation coefficient (R) close to 95% was obtained when both methods were used. While the province with the lowest standard deviation was Samsun, the highest was obtained in Diyarbakır.

Conclusion
This research proposes two distinct machine learning algorithms for one-day ahead RH estimates based on historical meteorological RH data spanning eight years. The findings of the research are as follows: machine learning techniques have been demonstrated to be beneficial in data forecasting, particularly in the prediction of RH levels. The ANFIS-FCM and LSTM have shown satisfactory estimate performance when using time series histories. Both models demonstrated adequate accuracy in the prediction of daily RH data, with MAE values ranging between 5.76 and 8.52%. When compared to the ANFIS-FCM model, the LSTM model estimates produced results that were closer to the observations in all provinces. The best results were found in Diyarbakır, where the dry climatic area predominates, using both the ANFIS-FCM and the LSTM approaches. The province of Samsun, on the other hand, had the lowest outcomes when it came to calculations. It was verified that the LSTM neural network, which was developed using a deep learning technique, is an extremely effective tool for short-term daily RH forecasting. The proposed time-series method based on ANFIS-FCM and LSTM neural networks provided predictions by taking into account the underlying periodicities in the data as it moves through time. In this sense, the most significant advantage of the univariate modeling approach proposed in this study is the ability to use data collected in the past as an input variable. As a result, the developed model can be used to forecast the one-day ahead RH value of any station without the need for complex calculations or extensive meteorological data. In future studies, it is aimed to realize models based on hybrid, new and innovative methods in order to obtain more accurate predictions in time series estimation of RH.