Analysis and Recommendation of Outdoor Activities for Smart City Users Based on Real-Time Contextual Data

1. Introduction

Transportation, infrastructure, and public services are just a few of the areas of urban life that smart cities are improving with the help of technology and real-time data. The availability of outdoor activities catered to the interests and requirements of its citizens is a critical component that can improve the quality of life in smart cities. The analysis’s goal is to investigate the possibility of using contextual data that are available in real time to suggest outside activities to smart city users. Smart city authorities may offer customized and pertinent recommendations for outdoor activities by utilizing data, including weather, air quality, crowd density, user preferences, and location-based information. This enhances the city’s general livability and sustainability in addition to fostering emotional and physical well-being. Utilizing machine learning for environmental monitoring entails utilizing computational methods to examine information gathered from diverse environmental sensors and sources. This method makes it possible to identify, forecast, and comprehend changes in the environment and their effects.

Many projects are being carried out in many nations to monitor pollutants as a result of the decreasing cost of pollution monitoring sensors. One example of a project is the installation of internet-connected pollutant measurement stations. Some of these programs have been gathering and archiving data for a number of years, which may contribute to our understanding of the issue of air pollution. Due to the importance of the air quality issue, several attempts have been made to forecast air quality using various techniques. Systems for predicting air quality can be used to characterize the issue and comprehend the connections between emissions, contaminants, weather, and other atmospheric variables. They can be used to inform air quality projections in the future. Deterministic models, which explain the atmospheric processes creating pollution with mathematical equations, are one type of forecasting approach.

Deep learning is a powerful tool. Applications of deep learning are numerous. The concept of Smart City is gaining momentum and has a lot of scope in the future. The idea of using deep learning for Smart City data analysis seemed intriguing and challenging. To be able to build a dynamic tool for analyzing data seemed like a perfect way to start exploring deep learning. Data are everywhere but not everyone knows how to make use of it. Neural networks help in processing information and provide accurate results. People are always into updating or rather adapting to newer technologies. There was once a time when everything had to be done manually, or in person. Be it ticket-booking, restaurant reservations or bill payments. Everything is online now, thanks to technology. Adapting to this way of life also has inspired us to build a project that can help people make certain decisions based on data analysis. This work is a result of inspiration and interest. It is designed to analyze the past data values of temperature and pollution levels and suggest whether the day is suitable for any outdoor activity that people plan for. An essential component of creating smart cities that are effective and sustainable is waste management. It is harder and harder to manage the trash efficiently as cities get bigger and more people live there. Smart technologies, on the other hand, provide creative ways to streamline the procedures involved in trash disposal, recycling, and collecting. To know more about deep learning and to be able to create something that helps people were the main motives behind it. This work tends to develop an application which helps in pollution and temperature prediction for outdoor activities using the Smart City dataset. This research will consist of a temperature-predicting model, a pollution-level predicting model and an analyzing model, which integrates to provide suggestions about people’s outdoor plans for a particular day. The training and testing data for air pollution prediction model and for the temperature prediction model was collected from the Beijing dataset¹ and Jena dataset,² used after a few modifications to fit our project requirements. The predictions made will be based on the trained data. Analyzing smart city data is useful in various fields. This model is trying to focus on using this analysis to plan outdoor activities based on environmental conditions. Given the data of the past few days, predicting the present data and analyzing it further to plan outdoor activities is what this research aims to achieve.

The rest of the paper first discusses the background on smart city, and forecasting techniques in Sec. 2. Thereafter author discusses the proposed temperature and pollution predictive model. This is followed by result and discussion in Sec. 4. Finally, conclusions are briefly outlined in Sec. 5.

2. Literature Review

A review of various approaches available for time-series prediction for forecasting the temperature and pollution values, certain deep learning concepts in terms of methodology, and a comparison between artificial neural network (ANN) and recurrent neural network (RNN), was carried out.

Interesting research articles that talk about using RNNs for time-series prediction³ gave an insight into how time-series prediction was carried out. Another research paper on House price prediction also helps us understand more about using long short-term memory (LSTM) networks for the prediction of values based on previous history. Certain articles from technical blogs also helped us gain more knowledge about LSTM and GRU networks and their usage.^4,5 The suggested model explores the utilization of bidirectional gated recurrent unit (GRU), a distinct type of RNN, along with 1D convents in experiments for forecasting PM2.5 concentrations in time series. The results are compared against conventional deep learning and machine learning models. The results illustrate the potential adequacy and competitiveness of the proposed method in forecasting PM2.5-time series data. Specifically, the predictive performance was superior when employing deep learning approaches compared to shallower machine learning models such as decision tree regression (DTR), support vector regression (SVR), and gradient boosted regression (GBR).⁶

In real time, the use of geographic information systems (GIS) in smart cities has grown significantly, and this Special Section has only touched the surface of the subject. Both computational power and digital data storage capacity are increasing, and in the big data era, real-time data acquisition, distribution, and processing are now feasible.⁷ Unlike previous approaches that required varying degrees of data assimilation or hybrid models, this approach is purely data-driven and can generate cheap, reliable forecasts for short durations. It can predict weather features quickly. A total of two models were trained and used to predict air temperature and relative humidity.⁸

In forecasting urban air pollutant PM2.5 levels, a hybrid model combining a convolutional neural network (CNN) and LSTM was proposed. A feature selection process involved choosing variables with higher correlation coefficients with PM2.5, including weather data and correlations with other monitoring stations. Subsequently, the suggested hybrid model was employed for prediction.⁹ Predicting the sub-daily temperature evolution involves employing video prediction models initially designed for computer vision applications. The results suggest that the performance of the model can be significantly improved by incorporating more advanced architectures such as the secure audio video profile (SAVP) model. Additionally, the inclusion of informative predictors, like the temperature at 850hPa, also contributes to enhancing the model’s performance.¹⁰

The research contributes by integrating LSTM through the optimization of its hyperparameters using a multiverse optimization metaheuristic algorithm. This approach is applied to predict NO₂ and SO₂ levels. The method is evaluated using an actual dataset of air pollution related to manufacturing from a Combined Cycle Power Plant.¹¹ Gaining proficiency in utilizing state-of-the-art technologies to create an affordable, energy-efficient, and precise real-time air pollution monitoring system is essential. Advances in next-generation air pollution monitoring systems for smart cities have been achieved through the integration of cloud computing, wireless sensor networks (WSN), low-power wide-area networks (LPWAN), the internet of things (IoT), micro-electro-mechanical systems (MEMS), and advanced sensing technologies.¹²

The involvement of smart tourists in the concept of a Smart Destination highlights substantial implications, emphasizing the need for strategic planning and robust public engagement in technological development. Successful, long-term outcomes necessitate a proactive approach to implementing smart technologies.¹³ Exploring the feasibility of utilizing existing public Wi-Fi infrastructure to observe crowds in unregulated, real-world environments. The monitoring system is easily accessible online and does not entail additional hardware investment or maintenance. This system also holds the potential to bolster public safety and management while enhancing the overall quality of public spaces.¹⁴ Leveraging context-aware features to create applications for smart cities, focusing on areas such as public health, tourism experiences, urban mobility, active citizenship, shopping experiences, urban infrastructure management, public alerts, recommenders, and smart environments.¹⁵ This research gives useful information on the future of AI-based indoor air quality (IAQ) forecasting. Readers will benefit from a more integrated view of IAQ and associated AI techniques within the built environment.¹⁶ For more accurate prediction and improved performance against measurement noise and nonlinearity, air quality alerts are sent out on time. The toolbox known as fractional-order modelling and control, or fractional-order modeling and control (FOMCON), is utilized to implement the estimation scheme.¹⁷ The assessing ability of machine learning techniques to forecast NO₂, SO₂, and PM10 in Amman, Jordan. We compared multiple machine learning methods like ANNs, SVR, DTR, and extreme gradient boosting. We also investigated the effect of the pollution station and the meteorological station distance on the prediction result as well as explored the most relevant seasonal variables and the most important minimal set of features required for prediction to improve the prediction time.¹⁸ The Gaussian Naive Bayes model achieves the highest accuracy and the Support Vector Machine model exhibits the lowest accuracy. XGBoost model performed best among all the other models and gets the highest linearity between the predicted and actual value.¹⁹ The developed model is quite useful for updating citizens about the predicted air quality of the urban spaces and protecting them from getting affected by poor ambient air quality. It can also be used to find the proper abatement strategies as well as operational measures.²⁰

Recurrent Neural Networks: These networks work on a feedback mechanism. They have a memory of what was computed which affects the output considerably. For certain applications, this mechanism is very important. Be it predicting the next word in a sentence or time-series prediction, past history is very important. They contain loops enabling information persistence. These loops can be thought of as a series of the same network with information passed on from one network to its successor. Figure 1 shows an unrolled RNN.

Long Short-Term Memory Networks: LSTM networks are designed for long-term memory storage. Their default behavior is to maintain information from past computations for a long time. These networks are perfectly suitable for forecasting pollution levels every hour.

Gated Recurrent Unit Networks: GRU networks are improvised versions of LSTM networks. They have gated mechanisms that enable them to retain past input even if irrelevant or old, and use it as required while giving the output. The work makes use of GRU networks for the temperature forecasting model.

These above factors make RNNs suitable for the presented research, hence based on the literature survey, the author has used RNNs to train our models as shown in Table 1. Subsequently, for better understating author has created a literature review in tabular form as shown in Table 2.

3. Proposed Solution

The work involves the integration of three models. The input values from smart city sensors are fed into corresponding neural network modules, i.e., the temperature model and the air pollution model. The output obtained is further sent as input to the analysis model to get the final output that suggests suitable outdoor activities based on predicted temperature and pollution values. Figure 2 shows the model discussed.

The three models are mentioned briefly in the following:

The Temperature Prediction Model: Constructed with GRU networks, this model incorporates 5 inputs, 32 neurons in a single hidden layer, and a final output layer.

The Pollution-Level Prediction Model: Constructed with LSTM RRNs, this model comprises 8 inputs, 50 neurons in the hidden layer, and a single output layer.

The Analysis Model: Results from the above models are passed to this model. It is responsible for analyzing the processed data and providing a meaningful output.

These three models integrate to create a complete system for real-time contextual analysis of smart city data. Figure 2 represents these models.

3.1. Air pollution forecasting model

This model focuses on forecasting air pollution values by utilizing historical data as input. It adopts an LSTM-RNN architecture, featuring an input layer with eight neurons, a hidden layer with fifty neurons, and an output layer predicting the air pollution value. The optimization is performed using the Adam optimizer, and the loss function employed is the root mean squared error (RMSE).

The data taken to train this model were from the Beijing, 2010 dataset, air pollution levels are measured in PM2.5 concentrations. The dataset has a total size of 2.23 Mb. Training and validation were done on the first years’ data, i.e., the model was trained on 8760 samples and validated on 35,039 samples, the rest of the data, i.e., the remaining three years’ data were used for testing.

Raw input data had to be normalized before it could train the model with it. The output of the model then had to be de-normalized to obtain actual results. Equations (1)–(3) represent the same.

Steps involved in normalizing the input data are as follows :

It was also achieved through sklearnfit_transform () function on the input data.

Steps involved in de-normalizing the output data are as follows :

Equations (4) and (5) represent the same. The model expects 3D data, hence the available 2D dataset had to be converted to 3D; the output hence obtained was converted back to 2D data to obtain the actual value.

Figure 3 shows the air pollution model.

3.2. Temperature forecasting model

The temperature values are forecasted by this model, leveraging past data as input. The model, designed as a GRU-RNN network, is composed of an input layer featuring five neurons, a hidden layer with 32 neurons, and an output layer with a single neuron predicting the temperature. The optimization is conducted using the RMSprop () optimizer, and the loss function employed is the mean absolute error (MAE).

The model was trained on Jena, 2020 dataset. Temperature values are in degree Celsius. The dataset has a total size of 54.9 Mb and due to this enormous size, the GRU network was chosen instead of LSTM. The network provides a faster way of training the data.

The data had to be modified and normalized before it was fed into the neural networks.

The same normalization function as discussed in Sec. 3.2 was used in this model as well. Figure 4 shows the temperature predictive model.

The Analysis Model

This model uses the outputs from the previous two models, processes it and provides suggestions based on pollution levels and temperature levels.

The temperature will be classified as hot, moderate and cold temperatures based on pre-determined thresholds. Pollution levels are classified as high and low based on PM2.5 concentrations. The set of activities suitable for a particular temperature or pollution value is suggested as the final output.

Error Functions

Error is calculated for each predicted value. Error functions used in the models are MAE and RMSE.

Mean Absolute Error

The MAE assesses the average magnitude of errors within a set of predictions, irrespective of their direction. It is calculated as the average across the test sample of the absolute differences between the predictions and the actual observations, with all individual differences assigned equal weight. The expression for MAE is shown in the following equation :

Example, the temperature model has a validation MAE of ∼0.26 and translates to an MAE of $\sim 2.4^{\circ }$C in the predicted values, as shown in Fig. 5.

Root Mean Squared Error

RMSE is a quadratic scoring rule that gauges the average magnitude of the error. It is calculated as the square root of the average of squared differences between predictions and actual observations.¹⁶ The expression for RMSE is shown in the following equation :

Example, the air pollution model achieves an RMSE of 27.2, which is lower than an RMSE of 30 found with a persistence model, as shown in Fig. 6.

Proposed Algorithm

The algorithm for the project involves the following steps:

Step 1: Modify the raw data by normalizing and fitting it as per the model’s requirements.

Step 2: Train the air pollution model and the temperature model with suitable training data.

Step 3: Validate the trained models using a set of validation data.

Step 4: Test the trained models using the testing data.

Step 5: Identify errors in the system, correct the abnormalities and test the models.

Step 6: Repeat Steps 4 and 5 to obtain optimized, efficient models.

Step 7: Feed the output values from the prediction models into the analysis model.

Step 8: Obtain the final output.

The Flowchart (See Fig. 7).

4. Result and Discussion

Air pollution levels prediction model (in PM2.5 concentration):

The graphs for the test cases are shown below:

Graph 1: Valid output — loss within the permissible range. Refer to Fig. 8.

Input values: [369, 354, 335, 394, 397, 395, 401, 406, 399, 373]

Predicted value: 391.32000

Actual value: 396

Graph 2: Invalid output — loss exceeded the permissible range of RMSE 27.17. Refer to Fig. 9.

Input values: [12, 15, 29, 44, 33, 22, 27, 26, 18, 27]

Predicted value: 25.45000

Actual value: 10

Graph 3: Valid output — loss within the permissible range. Refer to Fig. 10.

Input values: [12, 18, 15, 13, 12, 11, 15, 12, 15]

Predicted values: 13.69000

Actual values: 17

Graph 4: Valid output — loss within the permissible range. Refer to Fig. 11.

Input values: [160, 175, 166, 153, 165, 135, 270, 141, 137, 143]

Predicted value: 169.71000

Actual value: 167

Graph 5: Valid output — loss within the permissible range. Refer to Fig. 12.

Input values: [120, 124, 113, 104, 106, 121, 124, 135]

Predicted value: 119.74000

Actual value: 129

Graph 6: Valid output — loss within the permissible range. Refer to Fig. 13.

Input values: [96, 103, 138, 148, 126, 101, 104, 116, 143, 151]

Predicted value: 125.01000

Actual value: 149

X-axis contains input data, Y-axis contains the normalized values and the plot after the vertical line is the prediction.

The common observation is that, in all these graphs, the valid output graphs are closer to the original values that it is tested against. The invalid output graph has values that are a bit off from the original values.

The testing of air pollution model provided around 83% accurate, valid results and 17% inaccurate results, i.e., 5 out of 6 test cases were valid and one of them was invalid. Figure 14 gives the pictorial representation of the outcomes of tests carried out.

Temperature levels prediction model (in degree Celsius):

Graph 1: Valid output — loss within the permissible range. Refer to Fig. 15.

Input values: [21.60000, 21.50000, 21.58000, 22.22000, 22.05000, 21.87000, 21.98000, 21.96000, 21.98000, 21.94000]

Predicted value: 22.03800

Actual value: 22

Graph 2: Valid output — loss within the permissible range. Refer to Fig. 16.

Input values: [16.87000, 16.93000, 16.88000, 16.85000, 16.78000, 16.68000, 16.57000, 16.53000, 16.51000, 16.48000]

Predicted value: 16.72800

Actual value: 16.58000

Graph 3: Valid output — loss within the permissible range. Refer to Fig. 17.

Input values: [23.25000, 23.28000, 23.30000, 23.69000, 24.16000, 23.96000, 23.66000, 23.35000, 23.28000, 22.92000]

Predicted value: 23.74000

Actual value: 21.32000

Graph 4: Valid output — loss within the permissible range. Refer to Fig. 18.

Input values: [−2.59000, −2.89000, −3.22000, −4.00000, −4.45000, −4.09000, −3.76000, −3.93000, −4.05000, −3.35000, −3.16000]

Predicted value: −4.28000

Actual value: −4.23000

Graph 5: Valid output — loss within the permissible range. Refer to Fig. 19.

Input values: [11.51000, 11.72000, 12.06000, 12.60000, 13.90000, 14.47000, 14.93000, 15.24000, 15.13000, 15.41000]

Predicted value: 14.06800

Actual value: 15.87000

X-axis contains input data, Y-axis contains the normalized values and the plot after the vertical line is the prediction. The values obtained were accurate for all test cases, i.e., the author got 100% valid outputs. Figure 20 gives a pictorial representation of the outcome of tests carried out.

The models were successfully trained, validated and tested to predict accurate air pollution and temperature levels. These values were successfully analyzed by the analysis model and a final output consisting of a list of suggested outdoor activities was displayed.

As shown in Fig. 21, the air pollution model predicted 96.06700 PM2.5 concentration levels for the input values: [97.35461, 96.89057, 95.86968, 95.77687, 95.68407, 95.591255, 95.312836, 95.312836, 96.0553].

The graph contains plot of original (blue) values and predicted (orange) values. Loss incurred was within RMSE 27.17000. X-axis contains input data; Y-axis contains the normalized values and the plot after the vertical line is the prediction.

As shown in Fig. 22, the temperature model predicted 14.068^∘C for the input values: [11.51000, 11.72000, 12.06000, 12.6000, 13.9000, 14.47000, 14.93000, 15.24000, 15.13000, 15.41000] where the actual value is 15.87000.

The discussion is done on what privacy issues are brought up by the gathering and evaluation of contextual real-time data for suggested outdoor activities. How can user privacy be preserved while yet enabling efficient recommendation systems using privacy-preserving approaches like data reduction, encryption, and anonymization? What part may outdoor exercise suggestions play in encouraging urban dwellers’ physical and mental health as well as their social connectedness? In what ways may behavioral science knowledge be used in recommendation systems to promote habit development and positive behavior modification?

5. Conclusion

A common motive for any invention is to make people’s lives easier. Advancements in technologies are a potent tool to achieve this motive. There are significant opportunities to improve urban living conditions and encourage healthier lives through the analysis and suggestion of outdoor activities for users of smart cities based on contextual data collected in real time. Through the use of technological innovations, notably in the areas of IoT devices, machine learning, and data analytics, we can provide personalized and dynamic suggestions that address the varied requirements and inclinations of urban dwellers.

We can create sophisticated recommendation systems that take into account accessibility, inclusivity, and environmental sustainability in addition to suggesting relevant outdoor activities by integrating real-time contextual data sources like weather, air quality, traffic patterns, and user preferences. Furthermore, we can guarantee that suggestions for outdoor activities are in line with more general objectives of urban development and community ambitions by encouraging cooperation amongst stakeholders, including government agencies, urban planners, technology businesses, and community organizations. This cooperative strategy can result in the co-design and execution of creative solutions that improve smart city living standards while promoting environmental stewardship and social togetherness.

Deep learning which is an advanced machine learning approach, is proving to be very useful. The thought that computers can have neural networks and can almost think like humans is intriguing surely. Such a learning mechanism when applied to process big data results in interesting applications. One such application that the author has chosen is to provide a tool for planning outdoor events easily based on basic environmental factors.

The framework collects the real-time data and subsequently applies the current city context data and learning models like RNNs and CNNs, to analyze the data and predict real-time contextual information, including weather conditions, traffic patterns, and social events. The result shows that the proposed air pollution model predicted 96.067 PM2.5 concentration levels, subsequently the Temperature model predicted 14.068^∘C. Based on the model outcome people need not worry about unexpected weather or disappointment after reaching a place, due to environmental factors.

In conclusion, there is a great chance to employ technology to improve urban communities through the study and recommendation of outdoor activities for users of smart cities. Through the utilization of contextual data in real-time and cooperative collaborations, we can enable people to make well-informed decisions regarding their free time, resulting in urban settings that are more sustainable, happier, and healthier.

6. Limitations and Future Work

In order to collect real-time contextual data on users’ whereabouts, preferences, activities, and environmental conditions, future work may consider integrating IoT devices like wearable sensors, environmental monitors, and GPS trackers. Recommendations for outdoor activities should be integrated with smart city infrastructure, such as bike-sharing programs, smart parks, and public transit systems. Users may have easy access to resources and services that facilitate their preferred activities thanks to this connection. Utilize surveys, usability testing, and user feedback to continuously assess the efficacy and user satisfaction of outdoor activity suggestions. Over time, make use of this input to enhance the recommendation system’s relevance, accuracy, and usefulness. Develop systems and algorithms that dynamically modify suggested outdoor activities in response to shifting environmental factors and user limits, for instance, offering indoor activities when the air quality is low or proposing different paths for walking or cycling to avoid traffic or construction sites.

List of Abbreviations

ORCID

S. R. Mani Sekhar  https://orcid.org/0000-0003-3642-2021

D. M. Mushtaq Ahmed  https://orcid.org/0009-0003-9191-4245

G. M. Siddesh  https://orcid.org/0000-0002-6746-9319